In mid‑October 2026 I announced LFM2.5‑24B‑A2B, the largest member of the LFM2.5 family, positioned as an efficient local model for laptops and PCs. It uses a Mixture‑of‑Experts‑style architecture with only 2 B active parameters per token, giving fast inference while keeping the total parameter count high. The model ships with a 131 K‑token context window, a 128 K‑token vocabulary, and was pretrained on 33 T tokens.
LFM2.5‑24B‑A2B has 24 B total parameters.
It runs with 2 B active parameters per token.
The context window is 131 K tokens long.
The vocabulary contains 128 K tokens.
Pre‑training used 33 T tokens.
On 2026-10-13 I announced the release of LFM2.5-24B‑A2B, an extremely fast agentic model that we claim is competitive with models that have 50 % more parameters.
We just released LFM2.5-24B-A2B! 🥳
It's an extremely fast agentic model, competitive with models with 50% more parameters.
We expanded the context window to 131k and training up to 36T tokens.
On the post‑training side, we scaled up reinforcement learning and MOPD to include more environments and complex tasks.
You can test it today on Hugging Face or via our API, enjoy!
These decisions were driven by a desire to showcase a high‑throughput, cost‑effective architecture that matches larger models while keeping the parameter count at 24 B.