ByteDance opens Axion 1.0: Apache-2.0 agent weights + full recipe
A 35B/3B-active MoE and its entire ~18K-example recipe are downloadable now — matching Gemini 3.1 Pro on some science evals. Plus Claude Code locks down writes.

Copy markdown
The 35B checkpoint is live on Hugging Face
Axion 1.0-35B — a ~35B-total / 3B-active MoE on a Qwen3.5-MoE base, Apache 2.0, BF16 safetensors — is downloadable now and serves on vLLM or SGLang (eight GPUs, ~128K context). A larger 122B sibling is described in the paper but has no weights yet, and there's no hosted provider, so it's self-host from day one.
The real unlock: the whole recipe is open
ByteDance published the full pipeline — data construction, generation, filtering, verification, prompts and training code — plus the ~18.8K-row SFT dataset. The method is deliberately plain: two light SFT passes then RL, with no mid-training, DPO or on-policy distillation. It's a weekend-sized blueprint for training your own agent model.
3B active, near-Gemini science scores
The released checkpoint posts HLE 52.8 and FrontierScience Olympiad 85.0 — tying Gemini 3.1 Pro (85.0) on that eval and edging its HLE (51.4) while activating just 3B params. It's still well short of Claude Opus 5 (HLE 63.6), but that's frontier-adjacent reasoning on hardware you can actually rent.
Claude Code locks writes to your working dirs
v2.1.314 adds permissions.blockWritesOutsideWorkingDirectories, so Write, Edit and NotebookEdit refuse paths outside your working directories in every mode — a guardrail against agent write accidents. v2.1.318 also warns 10 days before your claude.ai sign-in expires, so long jobs don't die on a silent logout.