DeepSWE

Headlines · 4

1. 5d
   Meta releases Muse Wonder frontier model, Zuckerberg details $100B-plus bet

   Meta released the Muse Wonder frontier model on Oct. 13, a much larger successor to Muse Spark that Meta says leads on personal and professional agent tasks and in finance and legal evaluations, with strong coding results on DeepSWE, SWE-Marathon and Terminal-Bench 4, plus an improved "active memory" system tested on its internal Memory Bench. The model is available through Meta's model API and OpenRouter, with Muse Wonder Full reserved for Muse Code subscribers; Cursor also added it, reporting a 46.1% CursorBench score at default effort and pricing of $4 per million input and $20 per million output tokens. In a podcast interview, Zuckerberg and Meta Superintelligence Labs chief AI officer Alexandr Wang said running a 1 GW cluster costs about $15 billion a year with $50–60 billion in capex, that training a frontier model now runs into tens of billions of dollars, and that Meta committed $100–200 billion before seeing any model results.

   智东西 · 🔥 0
2. 7d
   Alibaba Qwen releases QwenGyre, an online RL framework for long-horizon agent tasks

   Alibaba's Qwen team released QwenGyre, an online reinforcement learning framework aimed at long-horizon agent tasks, on October 12. According to the company, such tasks can run for hours, involve hundreds of interdependent interactions with the environment and process millions of tokens; QwenGyre splits GPUs into compute units so that units freed when a batch of tasks finishes roll into training while running tasks keep their progress and caches, and it uses Token-in Token-out recording to organize shared-prefix trajectories into a tree, expanding only selected paths as training samples while Partial Scoring salvages feedback from timed-out but partially completed tasks. In the reported NL2RepoBench online RL experiment, with roughly 700,000 tokens per rollout, pass rate rose from 52.5% to 58.5% after 48 training steps, and QwenGyre delivered a 1.78x end-to-end speedup over an Async implementation at the same GPU budget, with cross-task end-to-end speedups on Qwen3.6-122B across NL2RepoBench, DeepSWE and TerminalBench.

   IT之家 AI · 🔥 0
3. Oct 10
   Xiaomi's MiMo-V2.6 report details $2.6M single-run RL post-training

   Xiaomi's MiMo team published a technical report, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," covering two sparse MoE models: MiMo-V2.6-Pro (1.02T total parameters, 42B activated) and MiMo-V2.6-Flash (310B total, 15B activated). The report says each RL step samples 1,568 prompts and roughly 25,000 agent trajectories totalling 2.7–3.7B tokens, with RL post-training costs of about $2.6M for Pro and $900K for Flash. The team also open-sourced more than 7,000 RL environments plus an end-to-end RL framework.

   InfoQ 中文 AI&大模型 · 🔥 0
4. Oct 8
   Step 5 Preview free for a week on OpenCode, Cline, Kilo and Nous Portal

   StepFun said its flagship model Step 5 Preview is free for one week in OpenCode, Cline, Kilo and Nous Research's Nous Portal for agentic and professional work. Material relayed by StepFun describes the model as a 600B-total, 27B-active MoE with a 1M-token context window and vision, and the OpenCode post also lists zero data retention. Nous Research is quoted as saying Step 5 Preview scored 33.89 on the Hermes Index, matching GPT-6 Luna, while Cline says it scores ahead of Kimi K3 and GLM-5.3 on DeepSWE; these benchmark comparisons are partner or vendor claims that have not been independently verified.

   阶跃星辰 StepFun · 🔥 0

Experience and discussion from the community

Share my DeepSWE experienceAsk about DeepSWE