Run fast, private agents on-device with LFM2.5-24B-A2B
- ID
- 34924
- Status
- summarized
- Published
- 14 Oct 2026, 1:05 AM
- Fetched
- 14 Oct 2026, 1:58 AM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 6.5
- Created
- 14 Oct 2026, 1:58 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_users
What happened
LiquidAI released LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with only 2B active parameters per token, positioned for running agents on your own hardware. It reports 184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU, and claims 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, a model with roughly 50% more total and active parameters. Versus its LFM2-24B-A2B predecessor it adds an explicit chain of thought, grows context from 32K to 131K tokens, doubles tokenizer vocabulary from 65K to 128K for non-Latin text, and scales pre-training from ~17T to ~33T tokens.
Why it matters
If you pay per-token for cloud agent loops, a 2B-active reasoning model means continuous multi-step tool-calling can run on a laptop or a single GPU instead of an API bill — the 131K context and 128K vocabulary are the parts that matter for long tool-use sessions and non-Latin prompts. But the LiveClaw 48.09-vs-38.19 comparison is LiquidAI's own benchmark, so run your own harness eval before replacing a cloud model in anything you ship.
Discussion angle
Which of your agent tasks actually need a large cloud model? Take one tool-calling workflow and compare a 2B-active local model at 131K context against what you currently call — then check whether the reported LiveClaw gap holds on your own task mix.