FireFlash: 300B terminal agent matches Fable 5 and GPT-6 Sol at 30x lower cost
Researchers published FireFlash, a 300B terminal and coding agent trained with full-parameter reinforcement learning on 64 NVIDIA B300 GPUs in under 5 days. It achieves 46.2% Pass@1 on Terminal-Bench 4.0, matching Fable 5 and GPT-6 Sol, with per-task cost over 30x lower than Fable 5 at highest inference intensity. The team released the complete training recipe.
Coverage timeline
机器之心机器之心
长程智能体（long-horizon agents）正从「修一个 bug」走向连续数小时的终端与编程工作，但它的强化学习训练也格外昂贵：前沿实验室可以在数千张 GPU 上扩展 RL，而对多数团队来说，一条训练样本往往就是一次长达 3 小时、上百轮工具调用的完整尝试，任何一个环节出错，浪费的都是成小时的 GPU 时间。 近日，Wenbo Hu、Nanyun Peng、Dakuo Wang、Kai-Wei Chang 等研究者在 Agent Economics 博客上发布了 FireFlash ：一个从 GLM-5.3-Flash 出发、经全参数强化学习训练得到的 300B 级终端与编程智能体。 FireFlash 在 Terminal-Bench 4.0 上的 Pass@1 达到 46.2% ，与 Fable 5、GPT-6 Sol 处于同一水平，而单个任务的成本比最高推理强度下的 Fable 5 低 30 倍以上 ；这一提升也迁移到了 DeepSWE v1.1 与 SWE-Bench Pro Verified 两个软件工程基准上。整个训练只用了 64 张 NVIDIA B300 、不到 5 天，团队同时公开了完整的训练配方。 博客标题：FireFlash: An Open Recipe for Long-Horizon Agentic RL Achieving the Performance-Cost Pareto Frontier 博客地址：https://agenteconomics.org/posts/agentic-rl-recipe 作者：Wenbo Hu、Jesse Cai、Po-nien Kung、Yuxuan Lu、Di Wu、Huimin Zeng、Nanyun Peng、Dakuo Wang、Kai-Wei Chang 图 1：Terminal-Bench 4.0 上 Pass@1 与平均单任务成本（对数坐标）的关系。虚线为 Pareto 前沿，箭头为 FireFlash 相对 GLM-5.3-Flash 的提升。由于模型由团队自行部署，FireFlash 按 GLM-5.3-Flash 的公开价格计费。 表 1：三个基准上的 Pass@1（5 次试验取平均）。GLM-5.3-Flash 为其在各基准上的官方成绩，FireFlash 在团队自己的
Related stories
vLLM: DeepSeek-V4.1-Flash on Vera Rubin NVL72 Hits 58x Throughput vs GB300
vLLM reported serving DeepSeek-V4.1-Flash on a single NVIDIA Vera Rubin NVL72 rack, achieving 1.7M tokens/s per GPU and a P90 of 387 tokens/s per user on SemiAnalysis AgentX. This represents 58x the throughput of a GB300 NVL72 at the same interactivity level.
1 reports · 1 sources
MiniMax unveils M3.1 Flash, claiming top agent scores and low cost
MiniMax released M3.1 Flash, the official version of the previously anonymous Space Bunny model, claiming a generational leap in agentic tasks, coding, computer use, and cybersecurity. The model scores 29.6% on Agents' Last Exam (ALE), the top for Chinese models, with pricing at 1 yuan per million input tokens and 4 yuan per million output tokens. Weights are due October 23, with day-0 support from vLLM on Blackwell, Rubin, and AMD MI350 GPUs.
Anthropic launches Claude Fable 5.2 and Mythos 5.2 with coding, reasoning, vision gains
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, flagship models aimed at complex technical work, with claimed improvements in coding, reasoning, vision, and communication. The models are priced at $8 per million input tokens and $40 per million output tokens, a 20% cut from 5.1, with a fast mode at $16/$80. During prerelease testing, the models formalized a proof of Fermat's Last Theorem in 30 hours, solved two open black hole physics problems, and scored 64.9% on the RoboDojo Sim benchmark, up from 27.6% for Fable 5.1.
14 reports · 7 sources916 points · 1051 comments
vLLM Serves DeepSeek-V4.1-Flash at 954K Tokens/s per GPU on GB300 NVL72
vLLM's blog reports that using P/D disaggregation and Wide-EP on a GB300 NVL72 rack, it serves DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 954K tokens/s per GPU, matching DeepInfer's performance. This result is 7.9x the aggregated single-node result, demonstrating significant scaling.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe