Fireworks AI

Fireworks AI demonstrates that GLM 5.1, when using Claude Opus 4.7 as a sparse advisor in the Legal Agent Benchmark, achieves 18/100 all-pass versus 14/100 for Opus alone at 39% of the cost.

FactoryAI通过微调Qwen模型，使Droid Shield在代码审查中比GPT-5.5多捕捉20%的真实秘密，同时降低成本和延迟。

Fireworks AI通过微调Qwen模型帮助LangChain实现评估成本降低100倍，替代GPT-5.5等闭源模型。

Fireworks AI 现已支持 DeepSeek V4 Flash 0731 模型的微调训练，提供 SFT/DPO/RL 三种训练方式，适用于编码代理和高并发场景，成本较传统方案降低 70%。

模型成本评估应关注任务完成成本而非每token价格，2400次测试表明按任务难度路由可同时优化成本与覆盖范围。

Fireworks AI与ClayRunHQ合作推动开放权重模型在GTM领域的落地，GLM-5.2和Kimi K2.6展现关键性能。

模型管理是SDLC扩展的真正瓶颈，Fireworks AI帮助FactoryAI实现模型管理效率提升。

Fireworks AI 现在支持对 GLM 5.2 进行微调，适用于代码生成任务，提升模型在特定代码库中的表现。

Fireworks AI 已解决前沿模型强化学习中的基础设施难题，实现训练与推理数值完全一致，并作为服务推出。

结合前沿封闭模型与开源模型可显著提升性能并降低成本，Fireworks AI 的研究验证了这一结论。

GLM 5.2 在性能和效率上表现优异，接近甚至超越 GPT 5.5 和 Opus 4.8。

Fireworks AI 已上线 GLM 5.2 模型，支持 1M-token 上下文，专注于代码生成，并在多个基准测试中表现优异。

Fireworks AI 现在支持 Kimi 2.7 的完全可训练，提供低成本构建竞争模型的方案。

Fireworks AI 提出通过缩短推理长度降低上下文开销，提升生成效率并减少成本。

Moonshot 推出 K2.7 Code 模型，推理 token 数减少 30%，同时在 Moonshot 编码基准中表现更优。

Step 3.7 Flash is a 196B MoE model designed from the ground up for inference efficiency, using MFA and AFD techniques to reduce KV-cache usage to ~22% of DeepSeek, supporting agent, coding, and multimodal workflows, open-sourced under Apache 2.0 and available on Fireworks.

开放权重模型在实际应用中表现出显著成本优势，RampLabs 使用 Kimi K2.6 和 DeepSeek V4 Pro 发现高危漏洞，成本降低 5 倍。

Fireworks AI provides a view that the success of a product lies in its uniqueness and adaptability as an RL environment.

Fireworks AI tests show baseline models had 20% retry rates in browser agent tasks, while Kimi K2.5/GLM-5/MiniMax M2.5 achieved near-zero retries with stable latency, directly impacting production system costs/delays/reliability.

Fireworks AI states that model fine-tuning now requires only a CLI command, 10 minutes of GPU time, and a few cents of compute cost, with full ownership of the weights. While 2026's off-the-shelf open models are sufficient for deployment, they remain just a starting point.

Fine-tuning bottlenecks aren’t about algorithms—they’re about integration complexity, slow iteration cycles, and unclear SFT vs RFT vs DPO choices. Fireworks AI enables secure, fast, and scalable tuning.

Fireworks AI introduces Step 3.7 Flash: a 198B sparse MoE VLM designed for inference from the start, with a 196B language backbone and 1.8B vision encoder, achieving up to 400 token/s on real-world agent workloads.

Routing and post-training open-source models significantly improve AI system accuracy, speed, and cost-efficiency. Harvey and Fireworks AI demonstrated that a hybrid architecture using GLM 5.1 as the primary worker with selective frontier model routing achieves superior quality and lower costs in legal tasks, proving this approach is a viable alternative to pure frontier models.

Fine-tuning on proprietary data is the most strategic AI advantage; prompts are easily copied, while models trained on private data are hard to replicate. OpenAI is restricting this path—companies must act now to retain SFT control.

Fireworks AI推出DeepSeek V4 Flash原生视觉模型，价格与文本版本相同但性能更优。

Fireworks AI团队使用GLM5.2 Fast模型在4天内完成原本预计1个月的工程任务，速度达400 tokens/秒，成本218美元。

Fireworks AI 每日处理 30 万亿 token，开放模型使用量持续增长，Factory AI 的开放模型使用量在过去一个月内增长了 3 倍。

The Cursor team achieved Composer 2.5 through reinforcement learning training rather than prompt engineering, with their large-scale RL program running inference on Fireworks, indicating that self-trained models will be the only way to maintain competitive moats after 2027.

Fireworks AI and Harvey team jointly launched the Legal Agent Benchmark to evaluate how open-weight models perform on long-horizon legal tasks.

Cursor AI launched Composer 2.5 on the Kimi K2.5 base model, achieving 85% performance gains from reinforcement learning, with Fireworks AI providing the RL infrastructure for scalable deployment.