Most LLMs trade off long-context reasoning for deployability; LFM2.5-24B-A2B flips that tradeoff by using MoE routing to keep per-token active parameters low, enabling continuous, tool-enabled agents on-device with very long contexts.
Key Capabilities
- Large but cheap per-token reasoning: 24B total parameters with ~2.45B active per token via MoE (64 experts, top-4 routing), which reduces compute and memory for each reasoning step. This enables sustained agent workloads on consumer hardware.
- Very long context and multilingual tokenization: 131,072-token context window and a 128k vocabulary (in-place tokenizer expansion) for better non-Latin handling and long multi-turn tool traces.
- Agentic post-training: multi-stage post-training including SFT, on-policy distillation, and agentic RL inside popular harnesses so the model learns function-calling, tool conventions, and multi-step interaction patterns.
- Deployment formats: native, FP8, GGUF, ONNX, MLX and framework support (Transformers, vLLM, SGLang, llama.cpp) for on-device CPU, Apple Silicon, and GPU serving.
Who it's for and tradeoffs
Great fit if you need an LLM that runs continuous, tool-using agents locally or at edge with long context needs and deterministic tool calling. Look elsewhere if your workload is purely knowledge-heavy (it’s recommended to augment with retrieval), or if you require a non-agentic, knowledge-first foundation model without the agent-oriented post-training biases. Benchmarks show strong agentic/tool-use performance versus comparable MoEs, but some knowledge benchmarks remain less dominant.
