---
title: "[AINews] not much happened today"
source_name: "Latent Space"
original_url: "https://www.latent.space/p/ainews-not-much-happened-today"
canonical_url: "https://www.traeai.com/articles/91345572-32d1-48fd-9001-1c0802ac017c"
content_type: "article"
language: "英文"
score: 7
tags: ["AI","深度学习","推理优化","开源模型"]
published_at: "2026-04-29T01:46:59+00:00"
created_at: "2026-04-29T03:44:16.514941+00:00"
---

# [AINews] not much happened today

Canonical URL: https://www.traeai.com/articles/91345572-32d1-48fd-9001-1c0802ac017c
Original source: https://www.latent.space/p/ainews-not-much-happened-today

## Summary

文章总结了近期AI领域的动态，包括vLLM 0.20更新、DeepSeek V4的推理优化及Poolside新模型发布。

## Key Takeaways

- vLLM 0.20提升内存效率和推理性能，支持多种硬件加速。
- DeepSeek V4在多硬件上优化显著，B300比H200快8倍。
- Poolside发布轻量化开源模型Laguna XS.2，适合单GPU部署。

## Content

Title: [AINews] not much happened today

URL Source: http://www.latent.space/p/ainews-not-much-happened-today

Published Time: 2026-04-29T01:46:59+00:00

Markdown Content:
# [AINews] not much happened today - Latent.Space

[![Image 1: Latent.Space](https://substackcdn.com/image/fetch/$s_!DbYa!,w_40,h_40,c_fill,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73b0838a-bd14-46a1-801c-b6a2046e5c1e_1130x1130.png)](http://www.latent.space/)

# [![Image 2: Latent.Space](https://substackcdn.com/image/fetch/$s_!1PJi!,e_trim:10:white/e_trim:10:transparent/h_72,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa4fe1182-38af-4a5d-bacc-439c36225e87_5000x1200.png)](http://www.latent.space/)

Subscribe Sign in

[AINews: Weekday Roundups](https://www.latent.space/s/ainews/?utm_source=substack&utm_medium=menu)

# [AINews] not much happened today

### a quiet day.

Apr 29, 2026

∙ Paid

18

Share

When we made the AINews → Substack move, we committed to having Matt Levine style op-eds every day, but some days there just isn’t much going on and we will just say so - we are working on small essays around inference demand and multiagents, but today is not that day.

Interesting model releases from [Nvidia Nemotron](https://x.com/NVIDIAAI/status/2049159441870717428?s=20), [Poolside](https://x.com/poolsideai/status/2049144111626670282?s=20), and [Alec Radford](https://x.com/status_effects/status/2048878495539843211), but it’s unclear any of them will stand the test of time. [GPT-6 hype](https://x.com/sama/status/2049241518540808440?s=20) is beginning.

> AI News for 4/27/2026-4/28/2026. 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!

* * *

# **AI Twitter Recap**

**Inference Systems, vLLM 0.20, and the Hardware/Kernel Race Around DeepSeek V4**

*   **vLLM’s latest release is heavily about memory and MoE serving efficiency**: [vLLM v0.20.0](https://x.com/TeksEdge/status/2048983564801450315) shipped with **TurboQuant 2-bit KV cache** for **4× KV capacity**, FA4 re-enabled for MLA prefill on **SM90+**, a new **vLLM IR** foundation, fused RMSNorm for a reported **2.1% end-to-end latency improvement**, plus support updates spanning **DeepSeek V4 MegaMoE on Blackwell**, Jetson Thor, ROCm, Intel XPU, and easier GB200/Grace-Blackwell setup. In parallel, [SemiAnalysis](https://x.com/SemiAnalysis_/status/2048957715955765284) highlighted early DeepSeek V4 Pro serving results on **B200/B300/H200/GB200 disaggregated setups**, claiming **B300 can be up to 8× faster than H200** for this workload and pointing to upcoming vLLM 0.20 benchmarking with **DeepGEMM MegaMoE**, which fuses **EP dispatch + EP combine + GEMMs + SwiGLU** into a single mega-kernel.

*   **DeepSeek support**: several posts focused on serving tradeoffs: [Jeremy Howard noted DeepSeek V4’s support for prefill](https://x.com/jeremyphoward/status/2049098509530583199) as a capability many providers have dropped, while [Maharshi](https://x.com/maharshii/status/2049058891389108640) pointed out the overheads of **dynamic activation quantization**, arguing that **static quantization** often wins on inference speed despite calibration cost. There was also growing interest in alternate stack portability: [teortaxesTex argued DeepSeek is structurally moving away from CUDA lock-in via TileKernels](https://x.com/teortaxesTex/status/2049185408785998217), suggesting model vendors may increasingly optimize for heterogeneous or domestic accelerator fleets rather than NVIDIA-only deployment.

**Open Model Releases: Poolside Laguna XS.2, NVIDIA Nemotron 3 Nano Omni, and TRELLIS.2**

*   **Poolside made its first public model release with an unusually deployment-friendly open-weight coder**: [@poolsideai announced Laguna XS.2](https://x.com/poolsideai/status/2049144111626670282), a **33B total / 3B active MoE** coding model trained fully in-house, released under **Apache 2.0**, and advertised as able to run on a **single GPU**. [Poolside’s broader release](https://x.com/eisokant/status/2049142230397370537) also included **Laguna M.1** and an agent harness, emphasizing that the company trained from scratch on its own **data, training infra, RL, and inference stack**. Community summaries added more color: [Aymeric Roucher](https://x.com/AymericRoucher/status/2049156715304935451) described two coder models—**225B/23B active** and **33B/3B active**—with **hybrid attention**, **FP8 KV cache**, and claimed performance near **Qwen-3.5**; [Ollama](https://x.com/ollama/status/2049184817603031463) shipped it immediately.

*   **NVIDIA’s Nemotron 3 Nano Omni was the day’s biggest infra-native model launch**: [@NVIDIAAI introduced Nemotron 3 Nano Omni](https://x.com/NVIDIAAI/status/2049159441870717428), an open **30B / A3B multimodal MoE** with **256K context** built for agentic workloads spanning **text, image, video, audio, and documents**. Distribution was immediate across the stack: [OpenRouter](https://x.com/OpenRouter/status/2049164366218772526), [LM Studio](https://x.com/lmstudio/status/2049172192705864091), [Ollama](https://x.com/ollama/status/2049194377751437470), [Unsloth](https://x.com/UnslothAI/status/2049161390150365344), [fal](https://x.com/fal/status/2049160999442198632), [Fireworks](https://x.com/FireworksAI_HQ/status/2049159136802398546), [DeepInfra](https://x.com/DeepInfra/status/2049158141070524815), [Together](https://x.com/togethercompute/status/2049160446708711883), [Baseten](https://x.com/baseten/status/2049160818575749300), [Canonical](https://x.com/Canonical/status/2049159988174602712), and others all announced same-day availability. Key specs surfaced in follow-on posts: [Piotr Żelasko](https://x.com/PiotrZelasko/status/2049162049599455725) described it as NVIDIA’s first **omni** release with speech/audio understanding backed by a **Parakeet encoder**, **English-only** for now, and a **5.95% WER** on the Open ASR leaderboard. Several hosts cited **~9× throughput** versus comparable open omni models.

*   **Other notable model/paper releases**: [Microsoft’s TRELLIS.2](https://x.com/kimmonismus/status/2049099376476459372) is an open-source **4B image-to-3D model** producing up to **1536³ PBR textured assets**, built on native 3D VAEs with **16× spatial compression**. On the world-model side, [World-R1](https://x.com/wjwang2003/status/2049136028968272260) claims existing video models already encode **3D structure** and can be “woken up” with **RL**, requiring **no architecture changes, no extra video training data, and no added inference cost**.

**Agents, Local-First Tooling, and Production Orchestration**

*   **Agent builders are shifting from demos to production primitives**: [Mistral launched Workflows in public preview](https://x.com/MistralAI/status/2049128071874179091) as an orchestration layer aimed at turning enterprise AI processes into durable, observable, fault-tolerant production systems. Related posts echoed the same theme: [Sydney Runkle framed durable execution](https://x.com/sydneyrunkle/status/2049132897227936073) as a key requirement for long-running agents, and [threepointone described work on subagents / agents-as-tools with persistence, streaming, and resumption](https://x.com/threepointone/status/2049088722835042475).

*   **Local/offline agents moved from aspiration to credible workflow**: [Teknium asserted “totally offline agents are possible”](https://x.com/Teknium/status/2048975223853350976), while [Niels Rogge demoed Pi + local models](https://x.com/NielsRogge/status/2049128153658839324) for desktop cleanup and [Google Gemma shared a tutorial for local coding agents](https://x.com/googlegemma/status/2049163687639007451). Hugging Face’s local push also showed up in adoption numbers: [Clement Delangue said 300,000 users have added hardware specs to the Hub](https://x.com/ClementDelangue/status/2049139562929143917) to discover what can run locally. Complementing this, [Ammaar open-sourced a vibe-coding app running Gemma 4 fully on-device with MLX](https://x.com/ammaar/status/2049169134429073471), and [Kimmonismus highlighted Sigma](https://x.com/kimmonismus/status/2049244932477759767), a private browser-based local-agent concept using open models.

*   **Hermes and adjacent agent harnesses are gaining real-world traction**: multiple posts reported Hermes outperforming OpenClaw in instruction-following or practical workflows, including [SecretArjun](https://x.com/SecretArjun/status/2049006382763110639), [somewheresy](https://x.com/somewheresy/status/2049089485938315614), and users deploying Hermes through [Telegram](https://x.com/lizliz404/status/2049084890717806877) or for [medical literature extraction](https://x.com/bobvarkey/status/2049120693649125687). On the research-agent side, [Hugging Face’s ML Intern](https://x.com/_lewtun/status/2049021398312468815) was trending among Spaces, and later gained [native metric logging + Trackio integration](https://x.com/akseljoonas/status/2049183527703396699) to make its training jobs observable rather than black-box.

**Benchmarks, Evals, and Research Findings Worth Watching**

*   **Model benchmarking remains fragmented, but a few signals stood out**: [Epoch reported GPT-5.5 Pro reaching](https://x.com/EpochAIResearch/status/2049186851844771888)**[159 on the Epoch Capabilities Index](https://x.com/EpochAIResearch/status/2049186851844771888)** and new highs on **FrontierMath**—**52% on Tiers 1–3** and **40% on Tier 4**—including two Tier 4 problems not previously solved by any model. Separately, [Greg Kamradt said ARC-AGI-3 testing for GPT-5.5 and Opus 4.7 had completed](https://x.com/GregKamradt/status/2049121093307547654), with failure modes now under analysis.

*   **Several new benchmarks target more realistic agent and engineering behavior**: [Lysandre announced a benchmark for making Transformers more agent-friendly](https://x.com/LysandreJik/status/2049053056814436352), and [VibeBench](https://x.com/jpschroeder/status/2049139723776495800) proposed subjective testing by **1,000 qualified software engineers** to measure how models actually feel in real work. On document intelligence, [LlamaIndex’s ParseBench](https://x.com/llama_index/status/2049139409316946011) emphasized that OCR benchmarks miss **semantic formatting** such as strikethroughs and superscripts, which materially alter meaning for agents.

*   **Research notes with concrete engineering implications**: [Rosinality flagged bugs in DeepSpeed and OpenRLHF that reduce SFT performance](https://x.com/rosinality/status/2049024030749970699), with implications for prior studies. [Arjun Kocher published a faithful implementation of Compressed Sparse Attention from the DeepSeek-V4 paper](https://x.com/arjunkocher/status/2049066844925936041). [che_shr_cat showed single-block transformers can solve Extreme Sudoku only with an explicit scratchpad and inverted routing init](https://x.com/che_shr_cat/status/2049081240762876261), otherwise performance is zero. On optimization, [Keller Jordan released a lightweight Modded-NanoGPT optimizer benchmark](https://x.com/kellerjordan0/status/2049193527440187494) designed to compare methods like **Muon** and **AdamW** on a reproducible speedrun-style task.

**Platform Economics, API Pricing, and Closed-Model Reliability Concerns**

## Keep reading with a 7-day free trial

Subscribe to Latent.Space to keep reading this post and get 7 days of free access to the full post archives.

[Start trial](https://www.latent.space/subscribe?simple=true&next=https%3A%2F%2Fwww.latent.space%2Fp%2Fainews-not-much-happened-today&utm_source=paywall-free-trial&utm_medium=web&utm_content=195821425&coupon=5fe099d9)

[Already a paid subscriber? **Sign in**](https://substack.com/sign-in?redirect=%2Fp%2Fainews-not-much-happened-today&for_pub=swyx&change_user=false)

Previous

© 2026 Latent.Space · [Privacy](https://substack.com/privacy) ∙ [Terms](https://substack.com/tos) ∙ [Collection notice](https://substack.com/ccpa#personal-data-collected)

[Start your Substack](https://substack.com/signup?utm_source=substack&utm_medium=web&utm_content=footer)[Get the app](https://substack.com/app/app-store-redirect?utm_campaign=app-marketing&utm_content=web-footer-button)

[Substack](https://substack.com/) is the home for great culture