GGUF Model: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF
Free Demo: https://playground.liquid.ai
“LFM2.5-24B-A2B is a reasoning model featuring 24B total parameters and just 2B active parameters, purpose-built for fast, private agents on edge devices. It serves as the largest and flagship capstone model of the LFM2.5 family, which now spans two orders of magnitude—ranging from LFM2.5-230M all the way to LFM2.5-24B-A2B.
- Best-in-Class Agent Performance: Matches Qwen3.6-35B-A3B—a model with ~50% more total and active parameters—on agent and tool-calling tasks, scoring 48.09 on the LiveClaw benchmark compared to Qwen’s 38.19.
- MoE-Speed Reasoning: Consistently “thinks before speaking.” Requiring only 2B active parameters per token keeps per-token inference costs exceptionally low.
- Energy-Efficient Inference: Runs up to 1.7x faster than comparable MoE models on edge devices, reaching 184 tok/s generation on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU.
LFM2.5-24B-A2B is a general-purpose text-only reasoning model with the following specifications:
- Total Parameters: 24.0B
- Active Parameters: 2.45B
- Layers: 40 layers (30 gated short-convolution blocks + 10 Grouped-Query Attention [GQA] blocks)
- Experts: 64 experts per MoE block, utilizing top-4 routing
- Training Compute Budget: 36 trillion tokens
- Vocabulary Size: 128,000
- Context Length: 131,072 tokens
- Supported Languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
- Recommended Generation Parameters:
- temperature: 0.1
- top_k: 50
- repetition_penalty: 1.05”
Top Discussions
Thrumpwart (⬆129): Wow, great achievement! Massive respect to any lab that openly publishes both their strengths and weaknesses. -Cubie- (⬆36): Agreed! Man, I just realized how rarely you ever see that nowadays. random-tomato (⬆23): Holy crap, a bar chart where the y-axis actually starts at 0? Respect. MiMillieuh (⬆53): Finally, some great news for us GPU-poor folks! Adventurous-Wing-988 (⬆26): Just finished testing it, and it’s incredible! I use Cursor for coding and a $6 GPT subscription for day-to-day chat. This model is more than capable of replacing that $6 subscription. Morphon (⬆28): Currently running it through my benchmark suite. So far, abstract reasoning and logic aren’t looking great—it failed all the humanities and social sciences tests. It seems its primary sweet spot is within agent harnesses, which I’ll be testing next. The good news is that on a 12GB 5070 + 64GB DDR3 (correction: DDR4) rig, it holds a steady sustained throughput of 64 t/s (Q4 quantization). On a Zen5 laptop, pure CPU inference is actually completely usable. So—it definitely takes the crown for speed. danigoncalves (⬆41): A faster Qwen 3.6 35B A3B? pmttyji (⬆34): It’s finally here! I was just about to ask in the comments, but found the answer right in the model card: LFM2.5-24B-A2B-DSpark, a draft model for speculative decoding in SGLang to achieve even faster generation, is coming soon.