---
title: "[AINews] DeepSeek V4 Pro (1.6T-A49B) and Flash (284B-A13B), Base and Instruct — runnable on Huawei Ascend chips"
source_name: "Latent Space"
original_url: "https://www.latent.space/p/ainews-deepseek-v4-pro-16t-a49b-and"
canonical_url: "https://www.traeai.com/articles/0ac6f0d1-2c76-473c-b348-b2dbeecb120f"
content_type: "article"
language: "英文"
score: 9.2
tags: ["DeepSeek","AI模型","华为","稀疏注意力"]
published_at: "2026-04-25T05:00:48+00:00"
created_at: "2026-04-25T05:59:29.643552+00:00"
---

# [AINews] DeepSeek V4 Pro (1.6T-A49B) and Flash (284B-A13B), Base and Instruct — runnable on Huawei Ascend chips

Canonical URL: https://www.traeai.com/articles/0ac6f0d1-2c76-473c-b348-b2dbeecb120f
Original source: https://www.latent.space/p/ainews-deepseek-v4-pro-16t-a49b-and

## Summary

DeepSeek 发布 V4 Pro 和 Flash 模型，支持 1M token 上下文，优化稀疏注意力技术，适配华为 Ascend 芯片，推动中国 AI 独立性。

## Key Takeaways

- DeepSeek V4 支持 1M token 上下文，显著提升长上下文处理能力。
- 采用 CSA/HCA 技术，相比 V3 减少 73% 的 FLOPs 和 90% 的 KV 缓存需求。
- 适配华为 Ascend 芯片，减少对 NVIDIA/CUDA 的依赖，具有重要地缘政治意义。

## Content

Title: [AINews] DeepSeek V4 Pro (1.6T-A49B) and Flash (284B-A13B), Base and Instruct — runnable on Huawei Ascend chips

URL Source: http://www.latent.space/p/ainews-deepseek-v4-pro-16t-a49b-and

Published Time: 2026-04-25T05:00:48+00:00

Markdown Content:
After a couple months’ delay and lots of speculation, [DeepSeek finally released the heavily anticipated DSV4](https://x.com/deepseek_ai/status/2047516922263285776?s=20), the first major version model since DSV3 (Dec 2024) and DSR1 (Jan 2025). It brings the DeepSeek family up in line with [Kimi K2.6](https://www.latent.space/p/ainews-moonshot-kimi-k26-the-worlds?utm_source=publication-search), the current open model leader, and [Xiaomi Mimo 2.5](https://x.com/ArtificialAnlys/status/2047799218828665093?s=20), a lesser known family [released 2 days ago](https://x.com/XiaomiMiMo/status/2046988157888209365?s=20).

[![Image 1](https://substackcdn.com/image/fetch/$s_!2kgW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa10f0270-c9c4-481b-962a-fcba50a2418b_1022x1104.png)](https://substackcdn.com/image/fetch/$s_!2kgW!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa10f0270-c9c4-481b-962a-fcba50a2418b_1022x1104.png)

The DSV4 family is roughly a Gemini 3.1, GPT 5.4, Opus 4.6 level model, up to 1.6T MOE withtrained on 32T tokens with [FP4](https://x.com/iscienceluvr/status/2047514399393579235?s=46), with 1M token context (supported by their new Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) techniques), and incredibly rarely, they released both the Base and Instruct versions - surely setting the stage for a possible “DeepSeek R2” in future, though this one already has reasoning effort.

[![Image 2](https://substackcdn.com/image/fetch/$s_!IADX!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff028c03e-53a7-4615-af85-fc5e6e11dab0_1226x940.png)](https://substackcdn.com/image/fetch/$s_!IADX!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff028c03e-53a7-4615-af85-fc5e6e11dab0_1226x940.png)

The [technical report](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf) is a typically dense 58 pages, demonstrating training and inference insights and improvements from [the Manifold Constrained Hyper-Connections (mHC) paper](https://arxiv.org/pdf/2512.24880) they released in January, continued usage of [Moonshot’s Muon](https://news.smol.ai/frozen-issues/25-07-11-kimi-k2.html), and CSA/HCA’s overall INCREDIBLE efficiency improvements on [DeepSeek 3.2-Exp’s already impressive Sparse Attention](https://news.smol.ai/frozen-issues/25-12-01-deepseek-32.html) - at 1M tokens, requiring only 27% of FLOPs and 10% of KV cache memory compared with DeepSeek-V3.2:

[![Image 3](https://substackcdn.com/image/fetch/$s_!ICSA!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff73baf75-34a0-46e8-8452-7cccd7481ba9_1156x730.png)](https://substackcdn.com/image/fetch/$s_!ICSA!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff73baf75-34a0-46e8-8452-7cccd7481ba9_1156x730.png)

The geopolitical backdrop behind the [Huawei CANN compatibility](https://x.com/jukan05/status/2047823601462812932) is DeepSeek weaning dependence off export-controlled NVIDIA/CUDA chips —Ascends are still [a quarter the supply](https://x.com/PalwinderCFA/status/2047614823102619974) of H100s, but this is an important milestone for Chinese total independence.

> AI News for 4/23/2026-4/24/2026. We checked 12 subreddits, [544 Twitters](https://twitter.com/i/lists/1585430245762441216) and no further Discords. [AINews’ website](https://news.smol.ai/) lets you search all past issues. As a reminder, [AINews is now a section of Latent Space](https://www.latent.space/p/2026). You can [opt in/out](https://support.substack.com/hc/en-us/articles/8914938285204-How-do-I-subscribe-to-or-unsubscribe-from-a-section-on-Substack) of email frequencies!

**Top Story: DeepSeek V4**

DeepSeek released **DeepSeek-V4 Pro** and **DeepSeek-V4 Flash**, its first major architecture refresh since V3 and first clear two-tier lineup, with **1M-token context**, hybrid reasoning/non-reasoning modes, an **MIT license**, and a technical report detailed enough that multiple researchers called it one of the most important or best-written model papers of the year. Across the reactions, the factual consensus is that V4 materially advances open-weight long-context and agentic coding performance while remaining somewhat behind the top closed frontier models overall. Independent benchmarkers place **V4 Pro around the #2 open-weights tier**, roughly near **Kimi K2.6 / GLM-5.1 / strong Claude Sonnet-class to Opus-ish** depending on benchmark and mode, with especially strong long-context and agentic performance; opinions diverge on how close it is to GPT-5.x / Opus 4.7 and on whether this is “democratizing” progress or an architecture so complex that few open labs can realistically reproduce it. Key sources include deep-dive commentary from [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953), [@scaling01](https://x.com/scaling01/status/2047618271310926151), [@nrehiew_](https://x.com/nrehiew_/status/2047665987730993363), [@ben_burtenshaw](https://x.com/ben_burtenshaw/status/2047646980139016560), [@TheZachMueller](https://x.com/TheZachMueller/status/2047702488418030066), [@ZhihuFrontier](https://x.com/ZhihuFrontier/status/2047664976215839021), and infra/vendor posts from [@vllm_project](https://x.com/vllm_project/status/2047843293447500069), [@NVIDIAAI](https://x.com/NVIDIAAI/status/2047765637808664759), and [@Togethercompute](https://x.com/togethercompute/status/2047743446522224987).

The most concrete technical claims repeated across the discussion:

*   **Two models**

    *   **V4 Pro:****1.6T total parameters / 49B active**

    *   **V4 Flash:****284B total / 13B active**

    *   Reported by [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953), [@teortaxesTex](https://x.com/teortaxesTex/status/2047630981364883816), [@baseten](https://x.com/baseten/status/2047779549644243146), [@NVIDIAAI](https://x.com/NVIDIAAI/status/2047765637808664759)

*   **Context**

    *   **1M tokens**, up from **128K in V3.2** per [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953)

    *   Multiple posters frame this as the headline achievement: “solid ultra-long context” [@teortaxesTex](https://x.com/teortaxesTex/status/2047623905754448043)

*   **Training scale**

    *   **32T–33T tokens** cited repeatedly

    *   [@nrehiew_](https://x.com/nrehiew_/status/2047666048334450754) notes **32T tokens** over **1.6T parameters**, i.e. roughly **20 tokens/parameter**

    *   [@teortaxesTex](https://x.com/teortaxesTex/status/2047630981364883816) cites **33T**

    *   [@nrehiew_](https://x.com/nrehiew_/status/2047840706874749076) estimates pretraining compute at **~1e25 FLOPs**

*   **Reasoning / modes**

    *   DeepSeek exposes **three reasoning modes** per [@Togethercompute](https://x.com/togethercompute/status/2047743446522224987)

    *   Hybrid “thinking/non-thinking” positioning noted by [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953)

*   **Long-context architecture**

    *   Several threads summarize a new hybrid attention system:

        *   shared KV vectors

        *   compressed KV streams

        *   sparse attention over compressed tokens

        *   local/sliding-window attention for nearby context

    *   [@ZhihuFrontier](https://x.com/ZhihuFrontier/status/2047664976215839021) gives the most compact public summary:

        *   **2× KV reduction** via shared key-value vectors

        *   **c4a ≈ 4× compression**

        *   **c128a ≈ 128× compression**

        *   **top-k sparse attention** on compressed tokens

        *   **128-token sliding window**

        *   **1M context KV cache = 9.62 GiB/sequence (bf16)**

        *   **8.7× smaller** than DeepSeek V3.2’s **83.9 GiB**

        *   FP4 index cache + FP8 attention cache gives another ~**2×** reduction

    *   [@ben_burtenshaw](https://x.com/ben_burtenshaw/status/2047646980139016560) condenses this to “**10× smaller KV cache**”

    *   [@TheZachMueller](https://x.com/TheZachMueller/status/2047702488418030066) and [@TheZachMueller](https://x.com/TheZachMueller/status/2047702996524405175) describe **CSA + HCA** layer patterns, with alternating layers and V4 Flash using sliding-window layers instead of HCA in some places

*   **Quantization / checkpoint format**

    *   [@LambdaAPI](https://x.com/LambdaAPI/status/2047654086263320965): checkpoint is **mixed FP4 + FP8**

        *   **MoE expert weights in FP4**

        *   attention / norm / router in **FP8**

        *   claim: the full model fits on a single **8×B200** node

*   **Inference hardware / serving**

    *   [@NVIDIAAI](https://x.com/NVIDIAAI/status/2047765637808664759): on **Blackwell Ultra**, V4 Pro can deliver **150+ TPS/user interactivity** for agentic workflows

    *   [@NVIDIAAI](https://x.com/NVIDIAAI/status/2047823093578518758): published day-0 V4 Pro performance pareto using **vLLM**

    *   [@SemiAnalysis_](https://x.com/SemiAnalysis_/status/2047726025748930687): day-0 support and benchmarking across **H200, MI355, B200, B300, GB200/300**

    *   [@Prince_Canuma](https://x.com/Prince_Canuma/status/2047685898163147125): **DeepSeek4-Flash on 256GB Mac**

    *   [@Prince_Canuma](https://x.com/Prince_Canuma/status/2047847095466385899): MLX quants published

    *   [@simonw](https://x.com/simonw/status/2047844236142497850) asks about smaller-RAM Mac viability, implying community interest but incomplete support story

    *   [@QuixiAI](https://x.com/QuixiAI/status/2047765475937890474) reminds users that many local stacks still lack tensor parallel, relevant because V4-class models strongly stress inference infra

*   **License / availability / pricing**

    *   **MIT license** per [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953)

    *   first-party API plus rapid third-party availability via [@Togethercompute](https://x.com/togethercompute/status/2047743446522224987), [@baseten](https://x.com/baseten/status/2047779549644243146), [@NousResearch](https://x.com/mr_r0b0t/status/2047673600900010044), [@Teknium](https://x.com/Teknium/status/2047798102091067677)

    *   **V4 Pro pricing:****$1.74 / $3.48 per 1M input/output tokens**

    *   **V4 Flash pricing:****$0.14 / $0.28**

    *   cache-hit pricing also given by [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953)

    *   [@scaling01](https://x.com/scaling01/status/2047707820552831028) views the pricing as a glimpse of future “Mythos-level” cheap coding models

    *   Reuters-via-posted quote from [@scaling01](https://x.com/scaling01/status/2047760776769720360): DeepSeek said **Pro pricing could fall sharply once Huawei Ascend 950 supernodes are deployed at scale in H2**

The most useful independent benchmark synthesis came from [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953):

*   **V4 Pro Max**: **52** on Artificial Analysis Intelligence Index

    *   up **10 points** from **V3.2 at 42**

    *   becomes **#2 open weights reasoning model**, behind **Kimi K2.6 (54)**

*   **V4 Flash Max**: **47**

    *   positioned around strong mid/high open models, “Claude Sonnet 4.6 max level intelligence”

*   **GDPval-AA** (agentic real-world work):

    *   **V4 Pro: 1554**, leading open-weight models

    *   ahead of **Kimi K2.6 (1484)**, **GLM-5.1 (1535)**, **MiniMax-M2.7 (1514)**

*   **AA-Omniscience**

    *   **V4 Pro: -10**, an 11-point improvement over V3.2

    *   but still paired with **94% hallucination rate**

    *   **V4 Flash: 96% hallucination rate**

*   **Cost to run AA Index**

    *   **V4 Pro: $1,071**

    *   **V4 Flash: $113**

*   **Output tokens used on AA Index**

    *   **V4 Pro: 190M**

    *   **V4 Flash: 240M**

    *   This is a major caveat: cheap per-token pricing does not imply cheap total task cost if the model spills huge token volumes

Additional eval perspectives:

*   [@arena](https://x.com/arena/status/2047714237502677405):

    *   **#2 open** in Text Arena overall at debut

    *   category wins/placements:

        *   **#1 Medical & Healthcare**

        *   **#15 Creative Writing**

        *   **#18 Multi-Turn**

    *   thinking variant:

        *   **#8 Math**

        *   **#9 Life/Physical/Social Science**

*   [@arena](https://x.com/arena/status/2047774037204742255) emphasizes the **Pro vs Flash tradeoff**:

    *   Pro ranks ~**30 places higher**

    *   costs **12× more**

    *   Flash is still competitive in Chinese, medicine, math

*   [@scaling01](https://x.com/scaling01/status/2047682465624445015):

    *   “~**Opus 4.5 estimate** holds for now, at least on SimpleBench”

*   [@scaling01](https://x.com/scaling01/status/2047733998714052819):

    *   V4 is “definitely better than GLM-5.1 but not quite Opus 4.7, GPT-5.4 or Gemini 3.1 Pro”

*   [@scaling01](https://x.com/scaling01/status/2047686712051048598) lists what scores would confirm <6 month gap:

    *   ARC-AGI-1 ~**75%**

    *   ARC-AGI-2 ~**35%**

    *   GSO ~**26%**

    *   METR **4.5–5 hours**

    *   WeirdML ~**63%**

*   [@TheZachMueller](https://x.com/TheZachMueller/status/2047719857869791352):

    *   on his evals, **Flash@max ≈ Pro@high on reasoning**

    *   Pro focuses more on knowledge (SimpleQA)

*   [@VictorTaelin](https://x.com/VictorTaelin/status/2047818978664268071):

    *   after fixing benchmark bugs and letting long-running models run longer, **DeepSeek and Kimi improved materially**

*   [@mbusigin](https://x.com/mbusigin/status/2047707082007220393):

    *   a simple negative early impression with no detail

*   [@petergostev](https://x.com/petergostev/status/2047773402090426548):

    *   on BullshitBench, not about capability but refusal/pushback behavior, GPT-5.5 underperformed; included here because many readers compare V4 in an eval-skeptical environment

*   V4 Pro / Flash were released with the specs above, **MIT-licensed**, **1M context**, and open technical documentation: [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953), [@TheZachMueller](https://x.com/TheZachMueller/status/2047626252425515240)

*   The architecture introduces a new long-context attention system with dramatic KV-cache reduction: [@ZhihuFrontier](https://x.com/ZhihuFrontier/status/2047664976215839021), [@ben_burtenshaw](https://x.com/ben_burtenshaw/status/2047646980139016560)

*   Independent benchmarkers broadly place V4 Pro near the very top of open weights but below the best proprietary models overall: [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953), [@arena](https://x.com/arena/status/2047714237502677405), [@scaling01](https://x.com/scaling01/status/2047733998714052819)

*   DeepSeek V4 is heavily token-intensive in some evaluations: [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953)

*   The checkpoint uses FP4/FP8 mixed precision and can fit on an 8×B200 node: [@LambdaAPI](https://x.com/LambdaAPI/status/2047654086263320965)

*   Rapid ecosystem support arrived via vLLM and other providers day 0: [@vllm_project](https://x.com/vllm_project/status/2047843293447500069), [@SemiAnalysis_](https://x.com/SemiAnalysis_/status/2047726025748930687)

*   “V4 is ~4–5 months behind the frontier” from [@scaling01](https://x.com/scaling01/status/2047618271310926151), [@scaling01](https://x.com/scaling01/status/2047622501241434581), [@scaling01](https://x.com/scaling01/status/2047626000091971811) is an informed estimate, not a measured fact

*   “Top three open” vs “only open model close to frontier” debate from [@teortaxesTex](https://x.com/teortaxesTex/status/2047616662879248828) is partly about benchmark trust and framing

*   “Strongest pretrained model we have” from [@teortaxesTex](https://x.com/teortaxesTex/status/2047630981364883816) is an opinion hinging on scale + architecture, not direct benchmark supremacy

*   “Most significant AI paper of the year” from [@Dorialexander](https://x.com/Dorialexander/status/2047632551326413109) is enthusiasm, not consensus

*   “This is what research should look like” from [@scaling01](https://x.com/scaling01/status/2047643722108579936) speaks to transparency/style rather than only capability

*   “Not exactly a democratizing technology” from [@teortaxesTex](https://x.com/teortaxesTex/status/2047840426371977467) is a strong architectural/political interpretation

**More favorable**

*   [@scaling01](https://x.com/scaling01/status/2047618271310926151): puts it at roughly **GPT-5.2 / Opus 4.5+ tier**

*   [@scaling01](https://x.com/scaling01/status/2047682465624445015): SimpleBench supports **~Opus 4.5**

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047630981364883816): argues it is the strongest pretraining base among opens and implies people are underestimating what post-training can do

**More skeptical**

*   [@scaling01](https://x.com/scaling01/status/2047733998714052819): below **Opus 4.7 / GPT-5.4 / Gemini 3.1 Pro**

*   [@scaling01](https://x.com/scaling01/status/2047622501241434581): the gap may widen again because closed labs have bigger models, better science/law/medicine coverage, faster inference with GB200s

*   [@mbusigin](https://x.com/mbusigin/status/2047707082007220393): early impressions “not great”

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047616897256947967): says polished models like **K2.6 and GLM 5.1** may still feel better in coding despite lower intrinsic capacity

A big split in reactions is that many technical readers think **the long-context architecture matters more than the raw benchmark position**.

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047623905754448043): “They’ve completed their quest: Solid Ultra-Long Context”

*   [@ben_burtenshaw](https://x.com/ben_burtenshaw/status/2047646980139016560): first open model where long context and agentic post-training “meet”

*   [@scaling01](https://x.com/scaling01/status/2047618271310926151): expects other open labs to adopt pieces of the architecture

*   [@Dorialexander](https://x.com/Dorialexander/status/2047632551326413109): frames Huawei/sovereignty constraints as an opportunity to reshape hardware and memory/interconnect design

*   [@jukan05](https://x.com/jukan05/status/2047861732702662741): reads the paper as evidence that NVIDIA’s hardware roadmap is unusually well aligned to where MoE/long-context models are going

This was one of the sharpest strategic disagreements.

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047840426371977467): says V4 is “not exactly a democratizing technology” because the architecture is too difficult for most labs to replicate

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047648219081974034): suggests even DeepSeek may not want to do this exact architecture again without refactoring

*   [@stochasticchasm](https://x.com/stochasticchasm/status/2047697372831183245): notes the sheer hyperparameter complexity is daunting

*   Against that, [@Prince_Canuma](https://x.com/Prince_Canuma/status/2047685898163147125) and [@Prince_Canuma](https://x.com/Prince_Canuma/status/2047847095466385899) show that the ecosystem is already compressing and adapting Flash for localish Apple Silicon use, softening the “not democratizing” claim on the inference side if not the training side

Several reactions suggest **Flash may be more important than Pro** for practical adoption.

*   [@arena](https://x.com/arena/status/2047774037204742255): Flash shifts the price/performance frontier

*   [@TheZachMueller](https://x.com/TheZachMueller/status/2047719857869791352): Flash@max ≈ Pro@high on reasoning tasks

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047864952862458009): benchmarks may underweight “legit 1M context for pennies”

*   [@Prince_Canuma](https://x.com/Prince_Canuma/status/2047685898163147125): Flash runs on **256GB Mac**

*   [@baseten](https://x.com/baseten/status/2047779549644243146) and [@Togethercompute](https://x.com/togethercompute/status/2047743446522224987) emphasize long-document analysis and agentic use cases where Flash’s economics matter

DeepSeek V4 was not discussed as a pure model release; it was treated as evidence in the larger US–China compute and sovereignty debate.

*   [@scaling01](https://x.com/scaling01/status/2047625331339661685): Chinese labs are already in or near “takeoff” in the sense that their models help build better models, though still shifted **5+ months** behind

*   [@scaling01](https://x.com/scaling01/status/2047622501241434581): thinks chip bans are likely to widen the gap in broad domains over time

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047608887616962992), [@teortaxesTex](https://x.com/teortaxesTex/status/2047631470664020211): disputes simplistic Huawei-dismissal and notes mixed Chinese sentiment toward Huawei

*   [@ogawa_tter](https://x.com/ogawa_tter/status/2047631993702363509): points to analysis of **Ascend 950** / A3 clusters and V4 deployment plans

*   [@Dorialexander](https://x.com/Dorialexander/status/2047632551326413109): argues the sovereignty play around Huawei may reshape hardware architecture

*   [@scaling01](https://x.com/scaling01/status/2047760776769720360): cites DeepSeek saying prices could drop sharply once **Ascend 950 supernodes** scale in H2

*   [@jukan05](https://x.com/jukan05/status/2047861732702662741): interprets V4 as validating NVIDIA’s Blackwell/Rubin/HBM/interconnect strategy

*   [@NVIDIAAI](https://x.com/NVIDIAAI/status/2047765637808664759), [@NVIDIAAI](https://x.com/NVIDIAAI/status/2047823093578518758): unsurprisingly highlight Blackwell day-0 performance, but this is vendor framing rather than independent proof of strategic superiority

There is also a more ideological thread:

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047645676234846459), [@teortaxesTex](https://x.com/teortaxesTex/status/2047638436295725080), [@teortaxesTex](https://x.com/teortaxesTex/status/2047835420755415472) argues that Western discourse often misreads Chinese labs as purely state proxies or distillation shops, and instead sees them as serious mission-driven actors. This is interpretive, but it helps explain why the release drew such emotionally charged geopolitical reactions.

A recurring undercurrent: does V4 mainly reflect architectural innovation, or can critics dismiss it as “distillation”?

*   [@yacineMTB](https://x.com/yacineMTB/status/2047628416514486661) speculates that some complaints about Chinese distillation may partly come from people discovering they’re outperformed

*   [@cloneofsimo](https://x.com/cloneofsimo/status/2047628636933812301): “Very interesting... given they distilled claude 🤔🤔”

*   [@kalomaze](https://x.com/kalomaze/status/2047762970931827125): jokes about DeepSeek training on DeepSeek reasoning traces

*   On the more substantive side, [@teortaxesTex](https://x.com/teortaxesTex/status/2047614729145745623) says DeepSeek’s writing quality, especially Chinese, reflects long-standing obsession with data cleanliness and cites job listings [@teortaxesTex](https://x.com/teortaxesTex/status/2047614852055683103), [@teortaxesTex](https://x.com/teortaxesTex/status/2047614975447855485)

*   [@nrehiew_](https://x.com/nrehiew_/status/2047666048334450754) notes the report still lacks much detail on pretraining data beyond standard categories

*   Overall, factual public evidence in this tweet set supports “DeepSeek trains at large scale with strong data work,” but not any strong claim about the degree of external distillation beyond speculation

Several researchers pointed out that V4 did not emerge from nowhere.

*   [@jaseweston](https://x.com/jaseweston/status/2047690308217926055): says DeepSeek uses **hash routing** from a 2021 ParlAI approach

*   [@suchenzang](https://x.com/suchenzang/status/2047772636881842629): criticizes routing-induced outliers, with a jab at hashing

*   [@teortaxesTex](https://x.com/teortaxesTex/status/2047844368883581404): notes Mixtral-style MoE was a reasonable earlier hack, but claims **DSMoE** changed things

*   [@art_zucker](https://x.com/art_zucker/status/2047619111082172548) broadly attacks MoEs as a dead end

*   [@gabriberton](https://x.com/gabriberton/status/2047835467551547587) counters that MoEs are provably effective despite inelegance

*   [@stochasticchasm](https://x.com/stochasticchasm/status/2047874903236645108) is even more positive: “MoEs are amazing”

This matters because V4 was read not just as a stronger checkpoint, but as a possible **new design point for open long-context MoEs**.

A striking amount of praise was directed not just at the model but at the paper/report quality.

*   [@scaling01](https://x.com/scaling01/status/2047618271310926151): “the technical paper is a big deal”

*   [@Dorialexander](https://x.com/Dorialexander/status/2047632551326413109): “most significant AI paper of the year”

*   [@morqon](https://x.com/morqon/status/2047643246923325833): “one of the best I’ve ever read”

*   [@scaling01](https://x.com/scaling01/status/2047643722108579936): “this is what research should look like”

*   [@TheZachMueller](https://x.com/TheZachMueller/status/2047626249116303561), [@iamgrigorev](https://x.com/iamgrigorev/status/2047641600591794546), [@nrehiew_](https://x.com/nrehiew_/status/2047665987730993363): all signal unusually high effort to digest and test the report

For expert readers, this is important because many frontier releases now arrive with sparse technical disclosure. V4’s report appears to have reset expectations for what a serious open release can look like.

Despite the enthusiasm, several caveats recur:

*   **Still behind closed frontier in aggregate capability**

    *   especially sciences/law/medicine and broad “general domains” per [@scaling01](https://x.com/scaling01/status/2047622501241434581)

*   **Reasoning RL may be undercooked**

    *   [@scaling01](https://x.com/scaling01/status/2047618271310926151): reasoning efficiency not much changed vs V3.2 Speciale

*   **Serving remains hard**

    *   [@scaling01](https://x.com/scaling01/status/2047643015859118167): many labs serve at only **20–30 tok/s** and limited concurrency; running evals can take a day

    *   [@ClementDelangue](https://x.com/ClementDelangue/status/2047664153439989823): acknowledges concurrency bottlenecks on HF

*   **High token usage**

    *   major practical caveat from [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953)

*   **API controls**

    *   [@stochasticchasm](https://x.com/stochasticchasm/status/2047717161070989499): notes DeepSeek API appears not to allow sampler control

*   **Adoptability**

    *   [@teortaxesTex](https://x.com/teortaxesTex/status/2047840426371977467): too complex for many labs to copy cleanly

Three implications stand out.

1.   **Open-weight long-context is no longer just marketing.**

V4’s strongest contribution may be proving that **1M context can be made operationally credible** in an open-weight model, with concrete KV-cache engineering and open inference support. This is why multiple posters focused less on benchmark deltas and more on systems design: [@ben_burtenshaw](https://x.com/ben_burtenshaw/status/2047646980139016560), [@ZhihuFrontier](https://x.com/ZhihuFrontier/status/2047664976215839021), [@scaling01](https://x.com/scaling01/status/2047618271310926151).

2.   **China’s top labs remain competitive in open models, even if not fully closing the closed-model gap.**

The benchmark picture across [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047735160544841953), [@arena](https://x.com/arena/status/2047714237502677405), and [@scaling01](https://x.com/scaling01/status/2047733998714052819) suggests Chinese labs now dominate much of the open-weight top tier: **Kimi, GLM, DeepSeek, and soon MiMo**.

3.   **The bar for “open” is rising from checkpoint release to full-stack co-design.**

V4 was instantly discussed alongside **vLLM**, **Blackwell**, **MLX quants**, **Mac viability**, **Ascend clusters**, and cache/memory architectures. In other words, “the model” is increasingly inseparable from the inference substrate.

**Infrastructure, inference, and local/open ecosystem**

*   Hugging Face launched **ML Intern**, an open-source CLI “AI intern” for ML work that can research papers, write code, run experiments, use HF datasets/jobs, search GitHub, and iterate up to **300 steps**, per [@MillieMarconnni](https://x.com/MillieMarconnni/status/2047639632859500691). Related sentiment: HF’s **$9 Pro** tier is unusually strong value per [@getpy](https://x.com/getpy/status/2047602009998794820).

*   Meta said it will add **tens of millions of AWS Graviton cores** to its compute portfolio to scale Meta AI and agentic systems for billions of users, per [@AIatMeta](https://x.com/AIatMeta/status/2047647617681957207).

*   Local/open coding stack momentum stayed strong:

    *   [@julien_c](https://x.com/julien_c/status/2047647522173104145): **Qwen3.6-27B via llama.cpp on a MacBook Pro** feels close to latest Opus for many coding tasks

    *   [@p0](https://x.com/p0/status/2047794814104862843): free CLI agent built with **Pi + Ollama + Gemma 4 + Parallel web search MCP**

    *   [@Prince_Canuma](https://x.com/Prince_Canuma/status/2047693737950670940): DeepSeek V4 quants incoming

    *   [@QuixiAI](https://x.com/QuixiAI/status/2047765475937890474): reminder that **llama.cpp / Ollama / LM Studio do not support tensor parallel**, pushing serious multi-GPU serving users toward **vLLM**

*   Nous/Hermes shipped heavily:

    *   Hermes Agent **v0.11.0** introduced a rewritten React TUI, dashboard plugin, theming, more inference providers, image backends, and QQBot support, per [@WesRoth](https://x.com/WesRoth/status/2047646749427216385)

    *   Hermes got broad praise and rapid support for both **DeepSeek V4** and **GPT-5.5**, via [@mr_r0b0t](https://x.com/mr_r0b0t/status/2047673600900010044), [@Teknium](https://x.com/Teknium/status/2047791512210293067)

    *   [@JulianGoldieSEO](https://x.com/JulianGoldieSEO/status/2047699587788361844) and [@LoicBerthelot](https://x.com/LoicBerthelot/status/2047690512199540959) compared Hermes favorably to OpenClaw on learning loops, memory, model support, deployment flexibility, and security

    *   A native Linux sandbox backend for Deep Agents using **bubblewrap + cgroups v2** was released by [@nu_b_kh](https://x.com/nu_b_kh/status/2047775326412136574)

**Research papers and benchmarks**

*   On-policy distillation token selection:

    *   [@TheTuringPost](https://x.com/TheTuringPost/status/2047617791709282405) highlights a paper showing only some tokens carry most learning signal; using **~50%** of tokens can match or beat full training and cut memory by **~47%**, while even **<10%** focused on confident-wrong tokens nearly matches full training.

*   Google Research pushed several ICLR demos:

    *   **MesaNet**, a transformer alternative / linear sequence layer optimized for in-context learning under fixed memory, via [@GoogleResearch](https://x.com/GoogleResearch/status/2047630714145776053)

    *   robotics/3D reasoning and efficient transformer work via [@GoogleResearch](https://x.com/GoogleResearch/status/2047675181808730197)

    *   “reasoning can lead to honesty” demo via [@GoogleResearch](https://x.com/GoogleResearch/status/2047704802163892576)

*   MIT **Hyperloop Transformers** mix looped and normal transformer blocks, using ~**50% fewer parameters** while beating regular transformers at **240M / 1B / 2B**, per [@TheTuringPost](https://x.com/TheTuringPost/status/2047720038342476187).

*   “Learning mechanics” tries to synthesize a theory of deep learning dynamics, via [@learning_mech](https://x.com/learning_mech/status/2047723849874330047).

*   Tool/agent systems papers:

    *   **Tool Attention Is All You Need** claims **95% tool-token reduction** (47.3k → 2.4k/turn) with dynamic gating and lazy schema loading, per [@omarsar0](https://x.com/omarsar0/status/2047725276851994639)

    *   **StructMem** for long-horizon structured memory highlighted by [@dair_ai](https://x.com/dair_ai/status/2047740873027543228)

    *   **HorizonBench** targets long-horizon personalization with shifting user preferences, via [@StellaLisy](https://x.com/StellaLisy/status/2047645651324821998)

*   Clarifying questions for software engineering:

    *   [@gneubig](https://x.com/gneubig/status/2047623214583492797) shared work on a model trained specifically to ask clarifying questions, improving results with fewer questions.

**GPT-5.5 rollout and coding agents**

*   OpenAI rolled **GPT-5.5** and **GPT-5.5 Pro** into API and ecosystem products with a **1M context window**, per [@OpenAI](https://x.com/OpenAI/status/2047743592278745425), [@OpenAIDevs](https://x.com/OpenAIDevs/status/2047742589982654915).

*   Distribution was immediate across Cursor, GitHub Copilot, Codex/OpenAI API, OpenRouter, Perplexity, Devin, Droid, Fleet, Deep Agents:

    *   [@cursor_ai](https://x.com/cursor_ai/status/2047744579127185843): GPT-5.5 is top on **CursorBench at 72.8%**

    *   [@cline](https://x.com/cline/status/2047769312514257148): **#1 on Terminal-Bench at 82.7**

    *   [@OpenAIDevs](https://x.com/OpenAIDevs/status/2047772632150675593): Perplexity Computer saw **56% fewer tokens** on complex tasks

    *   [@scaling01](https://x.com/scaling01/status/2047818395970904229): GPT-5.5 medium became strongest non-thinking model on LisanBench with **45.6% fewer tokens than GPT-5.4 medium** and higher scores

*   User feedback clustered around **better coding quality and token efficiency**, despite mixed feelings about some evals:

    *   [@almmaasoglu](https://x.com/almmaasoglu/status/2047745168141324559): best code they’ve read from an LLM; less verbose, less defensive

    *   [@KentonVarda](https://x.com/KentonVarda/status/2047788670728495142): caught a deep Cap’n Proto RPC corner case from a 6-year-old comment

    *   [@willdepue](https://x.com/willdepue/status/2047783399826292969): underwhelmed by evals, impressed in Codex on complex technical projects

    *   [@omarsar0](https://x.com/omarsar0/status/2047768166126809512): smooth switch from Claude Code to Codex/GPT-5.5 thanks to better “effort calibration”

*   Cursor also shipped **/multitask** async subagents and multi-root workspaces, via [@cursor_ai](https://x.com/cursor_ai/status/2047764651363180839).

*   There is growing market emphasis on **limits and economics** rather than tiny quality gaps:

    *   [@nrehiew_](https://x.com/nrehiew_/status/2047839351380537357) argues usage caps now matter more than small frontier deltas

    *   [@HamelHusain](https://x.com/HamelHusain/status/2047763070022479882) says Codex’s subscription structure makes it hard not to use

**Industry moves, funding, and policy**

*   Google reportedly plans to invest up to **$40B in Anthropic**, reported by [@FT](https://x.com/FT/status/2047715653553942997) and echoed by [@zerohedge](https://x.com/zerohedge/status/2047704883982180609). Reactions centered on how large Anthropic’s compute commitment may now be.

*   Cohere and Aleph Alpha announced a **Canada/Germany sovereign AI partnership**, framed as enterprise-grade and privacy/security focused by [@cohere](https://x.com/cohere/status/2047631725426000268), [@aidangomez](https://x.com/aidangomez/status/2047651054381052086), [@nickfrosst](https://x.com/nickfrosst/status/2047704679878996253#m).

*   ComfyUI raised **$30M at a $500M valuation**, while keeping core/open-local positioning, via [@yoland_yan](https://x.com/yoland_yan/status/2047731043000627263).

*   Mechanize announced **$9.1M** raised at a **$500M post-money valuation**, via [@MechanizeWork](https://x.com/MechanizeWork/status/2047732999878529037).

*   Arcee AI hired Cody Blakeney as Head of Research, emphasizing open-weight American frontier models, via [@code_star](https://x.com/code_star/status/2047765768658702467).

*   Safety / governance:

    *   OpenAI announced a **Bio Bug Bounty** for GPT-5.5, per [@OpenAINewsroom](https://x.com/OpenAINewsroom/status/2047670970526175310)

    *   Anthropic launched **Project Deal**, a marketplace where Claude negotiated on behalf of employees, and highlighted model-quality asymmetry and policy challenges, via [@AnthropicAI](https://x.com/AnthropicAI/status/2047728360818696302)

**Creative AI and multimodal**

*   GPT Image 2 + Seedance 2 workflows kept drawing attention:

    *   [@_OAK200](https://x.com/_OAK200/status/2047616640448078167) and [@awesome_visuals](https://x.com/awesome_visuals/status/2047609881104953658) showed high-fidelity image→video pipelines

    *   [@BoyuanChen0](https://x.com/BoyuanChen0/status/2047738501647728937) said **2K/4K** images are already available via experimental API and active fixes are underway

*   Kling announced native **4K output** and a **$25k** short film contest, via [@Kling_ai](https://x.com/Kling_ai/status/2047676942317678879).

*   Some evaluative nuance:

    *   [@goodside](https://x.com/goodside/status/2047728776520298646) noted GPT Images 2.0 could render a valid-looking Rubik’s Cube state, which is surprisingly hard

    *   [@venturetwins](https://x.com/venturetwins/status/2047820435543437630) framed recent image/video gains as a major step toward personalized game-like content generation

*   **[Deepseek V4 AGI comfirmed](https://www.reddit.com/r/LocalLLaMA/comments/1suolda/deepseek_v4_agi_comfirmed/)** (Activity: 1138): **The image is a meme and does not contain any technical content. The title “Deepseek V4 AGI confirmed” suggests a humorous or exaggerated claim about an AI model, possibly referencing advancements in artificial general intelligence (AGI). The comments further imply a satirical tone, mentioning uncensored datasets and military applications, which are likely not serious claims.** The comments reflect a satirical take on AI capabilities, with mentions of uncensored datasets and military applications, indicating skepticism or humor rather than a serious technical discussion.

    *   UserXtheUnknown discusses a test scenario with Deepseek V4, highlighting its tendency to overthink problems. The model interprets constraints like ‘using only one knife’ as mandatory rather than optional, which affects its problem-solving approach. This reflects a nuanced understanding of task constraints, but also indicates potential areas for improvement in handling implicit instructions.

*   **[Deepseek V4 Flash and Non-Flash Out on HuggingFace](https://www.reddit.com/r/LocalLLaMA/comments/1su3hdo/deepseek_v4_flash_and_nonflash_out_on_huggingface/)** (Activity: 1393): **DeepSeek V4 has been released on [HuggingFace](https://huggingface.co/collections/deepseek-ai/deepseek-v4), featuring two models: DeepSeek-V4-Pro with**`1.6T parameters`**(of which**`49B`**are activated) and DeepSeek-V4-Flash with**`284B parameters`**(with**`13B`**activated). Both models support a context length of**`one million tokens`**, which is significant for handling extensive sequences. The models are released under the MIT license, allowing for broad use and modification.** A notable comment highlights the challenge of hardware limitations, particularly RAM, when working with such large models. Another comment suggests the potential benefit of a `0.01bit quantization` to manage the model size more effectively.

    *   The DeepSeek-V4 models are notable for their massive parameter sizes, with the Pro version having 1.6 trillion parameters (49 billion activated) and the Flash version having 284 billion parameters (13 billion activated). Both models support an extensive context length of one million tokens, which is significant for handling large-scale data inputs and complex tasks.

    *   A user expressed interest in a 0.01-bit quantization of the DeepSeek-V4 models, which suggests a focus on reducing the model size and computational requirements while maintaining performance. Quantization is a common technique to optimize models for deployment on hardware with limited resources.

    *   The mention of the MIT license indicates that DeepSeek-V4 is open-source, allowing for broad use and modification by the community. This licensing choice can facilitate collaboration and innovation, as developers can freely integrate and adapt the models into their own projects.

*   **[Buried lede: Deepseek v4 Flash is incredibly inexpensive from the official API for its weight category](https://www.reddit.com/r/LocalLLaMA/comments/1su5gj5/buried_lede_deepseek_v4_flash_is_incredibly/)** (Activity: 404): **The image provides a comparison between two models, “deepseek-v4-flash” and “deepseek-v4-pro,” highlighting that the “deepseek-v4-flash” model is significantly more affordable in terms of input and output token costs. Despite its affordability, the model supports advanced features like JSON output, tool calls, and chat prefix completion in both non-thinking and thinking modes. The discussion around the image suggests that while the “deepseek-v4-flash” is marketed as inexpensive, some users argue that it is actually overpriced compared to previous versions when considering parameter scaling, with the “V3.2” model being cheaper per parameter.** Commenters discuss the impact of GPU shortages on current pricing, suggesting that prices may decrease as GPU production increases. There is also debate about the pricing strategy, with some users noting that the new model is more expensive per parameter compared to older versions.

    *   DistanceSolar1449 highlights a pricing comparison between DeepSeek V3.2 and V4 Flash, noting that V3.2 was priced at `$0.26/0.38` for input/output at `671b`, whereas V4 Flash is `$0.14/$0.28` at `284b`. This suggests that V4 Flash is actually more expensive if pricing were to scale linearly with parameters, challenging the notion of its cost-effectiveness.

    *   jwpbe provides a comparative analysis of DeepSeek V4 Flash’s API cost, stating that at `14 cents in / 28 cents out`, it is significantly cheaper than competitors like Minimax 2.7, which is `3x` the cost, and Qwen’s equivalent, which is even higher. They also mention that Trinity Thinking Large is twice as expensive, indicating that V4 Flash offers a competitive pricing advantage in the market.

    *   Worried-Squirrel2023 discusses the strategic implications of Huawei’s silicon developments, suggesting that DeepSeek’s pricing strategy involves trading NVIDIA margins for Ascend supply. They predict that once the `950 supernodes` scale, DeepSeek could potentially undercut competitors in the open weights tier, leveraging Huawei’s advancements to optimize costs.

*   **[Deepseek has released DeepEP V2 and TileKernels.](https://www.reddit.com/r/LocalLLaMA/comments/1ste9zs/deepseek_has_released_deepep_v2_and_tilekernels/)** (Activity: 396): **Deepseek has released DeepEP V2 and TileKernels, which are significant advancements in AI model optimization and parallelization. DeepEP V2 focuses on enhancing model efficiency and accuracy, while TileKernels introduces a novel parallelization technique that reportedly scales linearly, meaning that doubling computational capacity results in a doubling of processing speed. This release is open-sourced, fostering transparency and collaboration in AI research. For more details, see the [DeepEP V2 pull request](https://github.com/deepseek-ai/DeepEP/pull/605) and the [TileKernels repository](https://github.com/deepseek-ai/TileKernels).** One commenter highlights that **Deepseek** is fulfilling a role that **OpenAI** was expected to play by advancing research and sharing findings openly, which builds goodwill despite proprietary technologies. Another commenter questions if the parallelization technique indeed scales linearly, suggesting a significant technical breakthrough if true.

    *   **DeepEP V2 and TileKernels** by DeepSeek are noted for their potential advancements in parallelization techniques. A user speculates that these techniques might achieve linear scaling, meaning that doubling computational capacity could directly double processing speed. This could represent a significant efficiency improvement in model training and inference.

    *   There is speculation about DeepSeek’s hardware usage, particularly regarding the SM100 and Blackwell GPUs. One commenter suggests that DeepSeek might be using Blackwell GPUs for training, possibly through rented B200 units on Vast.ai. This hardware choice could influence the performance and capabilities of their models.

    *   The potential innovations in DeepSeek’s next model, possibly named v4, are highlighted. The focus is on the integration of Engram and mHC technologies, which are expected to play a crucial role in the model’s performance. The success of these innovations will likely depend on the new dataset DeepSeek has developed.

*   **[This is where we are right now, LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/comments/1suqfba/this_is_where_we_are_right_now_localllama/)** (Activity: 1755): **The image depicts a MacBook Pro running a Qwen3.6 27B model via Llama.cpp, showcasing the capability of executing complex AI models locally, even in airplane mode. This highlights the potential for local AI models to enhance efficiency, security, privacy, and sovereignty by operating independently of cloud services. The post underscores the technological advancement in making powerful AI models accessible on personal devices, emphasizing the importance of local execution for privacy and control.** Commenters express skepticism about the overstatement of the Qwen3.6-27B model’s capabilities, suggesting that while it is impressive for its size, it does not match the performance of more advanced models like Sonnet or Opus. There is concern that exaggerated claims could lead to user disappointment and backlash against the broader LLM community.

    *   **ttkciar** highlights the potential for user disappointment with the Qwen3.6-27B model, noting that while it’s impressive for its size and suitable for agentic code generation, it doesn’t match the capabilities of more advanced models like Sonnet or Opus. The concern is that overhyping its abilities could lead to backlash against the broader LLM community, not just the individual making the claims.

    *   **sooki10** agrees that while the model is impressive for local coding tasks, comparing it to more advanced models like Opus is misleading and could undermine the credibility of the claims being made. This suggests a need for more accurate benchmarking and communication about the model’s capabilities to manage user expectations effectively.

    *   **Melodic_Reality_646** points out the disparity in resources, comparing the use of a high-end 128GB RAM m5max system to a more accessible setup. This highlights the importance of considering hardware limitations when evaluating model performance, as not all users have access to such powerful systems, which can skew perceptions of a model’s capabilities.

*   **[DS4-Flash vs Qwen3.6](https://www.reddit.com/r/LocalLLaMA/comments/1sub71w/ds4flash_vs_qwen36/)** (Activity: 470): **The image presents a benchmark comparison between DS4-Flash Max and Qwen3.6 models, specifically the**`35B-A3B`**and**`27B`**versions. The chart highlights that DS4-Flash Max generally outperforms the Qwen models across various categories, particularly excelling in ‘LiveCodeBench’ and ‘HLE’ benchmarks. This suggests that DS4-Flash Max may have superior capabilities in coding and reasoning tasks. The discussion in the comments hints at the potential for larger models like a**`122B`**version of Qwen3.6, and emphasizes the significance of the**`1M token context`**feature, which could impact performance in other benchmarks like ‘omniscense’.** Commenters note that despite DS4-Flash Max’s larger size, its performance is only slightly better than Qwen3.6, raising questions about efficiency versus scale. The `1M token context` is highlighted as a significant feature that could influence future benchmark results.

    *   **Rascazzione** highlights the significant increase in context length with Qwen 3.6, noting its ability to handle a 1 million token context. This is a substantial improvement over previous models and could have significant implications for tasks requiring extensive context handling, such as document summarization or complex dialogue systems.

    *   **LinkSea8324** points out the size difference between the models, with DS4-Flash at 284 billion parameters compared to Qwen 3.6’s 27 billion. This raises questions about the efficiency and performance trade-offs between model size and capability, especially in terms of computational resources and inference speed.

    *   **madsheepPL** discusses the non-linear nature of benchmark improvements, suggesting that even if a model appears only slightly better in benchmarks, the practical implications can be more significant. They emphasize that improvements in scores are not directly proportional and can have varying impacts on real-world applications.

*   **[Qwen 3.6 27B Makes Huge Gains in Agency on Artificial Analysis - Ties with Sonnet 4.6](https://www.reddit.com/r/LocalLLaMA/comments/1strodp/qwen_36_27b_makes_huge_gains_in_agency_on/)** (Activity: 964): **Qwen 3.6 27B has achieved parity with Sonnet 4.6 on the Agentic Index from Artificial Analysis, surpassing models like Gemini 3.1 Pro Preview, GPT 5.2 and 5.3, and MiniMax 2.7. The model shows improvements across all indices, although the gains in the Coding Index are less pronounced due to its reliance on benchmarks like Terminal Bench Hard and SciCode, which are considered unconventional. The focus of training appears to be on agentic applications for OpenClaw/Hermes, highlighting the potential of smaller models to approach frontier capabilities. Anticipation is building for the upcoming Qwen 3.6 122B model.** Commenters express excitement about the potential of smaller models like Qwen 3.6 27B, noting the significant improvements and potential for future versions. However, there is skepticism about the extent of these gains, suggesting that some improvements might be due to ‘benchmaxxing’ rather than inherent model capabilities.

    *   Iory1998 highlights the impressive performance of the Qwen 3.6 27B model, noting that it surpasses a 670B model from the previous year. They mention running the Q8 version at 170K with KV cache at FP16 on an RTX 3090 and RTX 5070ti, utilizing 40GB of VRAM, which underscores the model’s efficiency and power.

    *   AngeloKappos discusses the narrowing benchmark gap, sharing their experience running the Qwen3-30b-a3b model on an M2 chip. They note its capability to handle multi-step tool calls effectively, suggesting that if the 27B dense model performs this well, the upcoming 122B model could pose challenges for API providers due to its potential performance.

    *   Velocita84 raises a point about potential “benchmaxxing” in the reported performance gains of the Qwen 3.6 27B model, implying that some of the improvements might be attributed to optimized benchmarking rather than inherent model capabilities. This suggests a need for scrutiny in evaluating model performance claims.

*   **[Compared QWEN 3.6 35B with QWEN 3.6 27B for coding primitives](https://www.reddit.com/r/LocalLLaMA/comments/1styxdy/compared_qwen_36_35b_with_qwen_36_27b_for_coding/)** (Activity: 491): **The post compares two versions of the QWEN 3.6 model, specifically the**`35B`**and**`27B`**parameter versions, on a MacBook Pro M5 MAX with**`64GB`**RAM. The**`35B`**model achieves**`72 TPS`**(tokens per second), while the**`27B`**model achieves**`18 TPS`**. Despite the slower speed, the**`27B`**model produces more precise and correct results for coding tasks, whereas the**`35B`**model is faster but less accurate. The test involved generating a single HTML file to simulate a moving car with a parallax effect, using no external libraries. The models were hosted using [Atomic.Chat](http://atomic.chat/), with source code available on [GitHub](https://github.com/AtomicBot-ai/Atomic-Chat).** One comment highlights the output of the `Qwen 3.6 27B FP8` model using opencode, taking approximately `52 seconds`. Another comment provides a visual comparison with the `Qwen 3.5 27B Q3` model, suggesting differences in output quality.

    *   The user ‘sacrelege’ shared a performance result for the Qwen 3.6 27B model using FP8 precision, noting that it took approximately 52 seconds to complete a task with ‘opencode’. This suggests a focus on optimizing model performance through precision adjustments, which can significantly impact computational efficiency and speed.

    *   User ‘nikhilprasanth’ provided a visual comparison for the Qwen 3.5 27B Q3 model, indicating a potential interest in comparing different versions and quantization levels of the Qwen models. This highlights the importance of understanding how different model configurations can affect performance and output quality.

    *   ‘Technical-Earth-3254’ inquired about the quantization methods used in the tests, which is crucial for understanding the trade-offs between model size, speed, and accuracy. Quantization can greatly influence the efficiency of large models like Qwen, especially in resource-constrained environments.

*   **[Qwen 3.6 27B is a BEAST](https://www.reddit.com/r/LocalLLaMA/comments/1steip4/qwen_36_27b_is_a_beast/)** (Activity: 1239): **The post discusses the performance of the Qwen 3.6 27B model on a high-end laptop with an RTX 5090 GPU and**`24GB VRAM`**, highlighting its effectiveness for pyspark/python and data transformation debugging tasks. The user employs llama.cpp with**`q4_k_m`**at**`q4_0`**and is exploring further optimizations with IQ4_XS at**`200k q8_0`**. The user has not yet implemented speculative decoding. The setup includes an ASUS ROG Strix SCAR 18 with**`64GB DDR5 RAM`**.** Comments suggest avoiding kv cache as q4 for coding, recommending `q8` for `130k` context. Another comment anticipates performance improvements with upcoming releases from **z-lab** and a specific [GitHub pull request](https://github.com/ggml-org/llama.cpp/pull/22105) that promises a `2x` decode speed increase. There is also curiosity about the model’s performance on systems with `16GB VRAM` and `32GB DDR5 RAM` with offloading.

    *   sagiroth highlights a technical consideration when using Qwen 3.6 27B for coding tasks, advising against using the KV cache as q4 due to limitations, and instead suggests using q8 to achieve a `130k` context window, which can significantly enhance performance for large context tasks.

    *   inkberk points out an upcoming improvement in decoding speed, referencing a pull request [#22105](https://github.com/ggml-org/llama.cpp/pull/22105) on the `llama.cpp` repository. This update, along with the anticipated release of the ‘dflash drafter’ by z-lab, promises a potential `2x` increase in decode speed, which could greatly benefit users in terms of efficiency.

    *   Johnny_Rell inquires about the performance of Qwen 3.6 27B on a system with `16 GB VRAM` and `32 GB DDR5`, specifically regarding the effectiveness of offloading. This suggests a focus on optimizing resource allocation to handle the model’s demands, which is crucial for running large models efficiently on consumer-grade hardware.