Liquid AI's open 24B LFM2.5 runs agents at 184 tok/s on a laptop
The 2B-active MoE beats Qwen's 35B on most tool-use tests and ships GGUF/MLX day one — plus Anthropic walls new API accounts off from editing Claude's context.
Copy markdown
Open weights, and it's built for agents
LFM2.5-24B-A2B is a 24B mixture-of-experts with just 2B active per token and a 131K context, downloadable from Hugging Face today. On Liquid's numbers the reasoning model tops Qwen3.6-35B-A3B on five of eight tool-use benchmarks — ToolSandbox 80.7, IFEval 96.8 — at a fraction of the active params.
184 tok/s on a MacBook, $0.10 to host
It decodes at 184 tok/s on an Apple M5 Max and 114 on a Ryzen AI Max+ 395, with GGUF, MLX, vLLM, SGLang and ONNX all supported on day one. The hosted API is $0.10/$1.00 per 1M tokens with $2 free to start — enough to stand up a local-first, tool-calling agent this weekend.
The catch: not OSI-open, and weak at code
Weights ship under Liquid's own license — free for individuals and small companies, paid above a revenue threshold — not Apache or MIT. And Qwen still wins on coding and math (SWE-Bench 51 vs 66), so reach for this for on-device tool use, not as your main coder.
Meanwhile: Anthropic locks down API prefill
Under new 'preserved thinking' rules, API accounts created on or after Aug 31 can no longer edit the context around a thinking block for Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 — a mismatch now returns an error (or opt into non-strict mode to drop the blocks). If you rewrite assistant turns on a newer account, that pattern just broke.