On Oct 13 Liquid AI released LFM2.5-24B-A2B, the largest and final LFM2.5 model: 24B total parameters, ~2.45B active per token, 131K context, and always-on reasoning. By Liquid's own numbers it leads or matches Qwen3.6-35B-A3B on 5 of 8 tool-use/agent benchmarks (AutomationBench 27.69 vs 8.72) while trailing on SWE-Bench Verified, GPQA and other knowledge/coding tests. Open weights are on Hugging Face with GGUF, MLX, FP8 and ONNX builds; API pricing is $0.10 / $1.00 per million tokens and it is also on OpenRouter.
Key Takeaways
- ✓Specs (model card): 24.0B total / 2.45B active, 40 layers (30 gated short-conv + 10 GQA), 64 experts top-4, 131,072 context, 128K vocab
- ✓Agent benchmarks (vendor-reported, vs Qwen3.6-35B-A3B): AutomationBench 27.69 vs 8.72, LiveClaw 48.09 vs 38.19, ToolSandbox 80.71 vs 75.02, BFCLv4 65.01 vs 65.00
- ✓Disclosed weaknesses: SWE-Bench Verified (OpenCode) 51.00 vs 66.20, GPQA Diamond 74.42 vs 83.69, WebArena Verified 26.75 vs 37.62
- ✓Speed: 184 tok/s decode on Apple M5 Max, 114 tok/s on Ryzen AI Max+ 395, 21K+ output tok/s on one B300 with SGLang
- ✓Training: pretraining scaled 17T to 33T tokens; agentic RL inside OpenClaw, Hermes Agent, Pi and OpenCode; API $0.10 in / $1.00 out per M tokens
-oOdUn61bDRSykk4H951SoDOHzqoOe1.png)
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Liquid AI's LFM2.5-24B-A2B (released Oct 13) is the largest and final LFM2.5 model and targets fast, private agents on consumer hardware. It keeps the LFM2-24B-A2B hybrid design (40 layers: 30 gated short-convolution blocks plus 10 GQA; 64 experts per MoE block with top-4 routing; 24B total, ~2.45B active), extends context from 32K to 131K, doubles the vocabulary to 128K via in-place tokenizer expansion, and becomes a reasoning-only model that always thinks first. Pretraining grew from 17T to 33T tokens, mid-training added 2T tokens at 32K and 300B at 131K, and post-training ends with agentic RL inside OpenClaw, Hermes Agent, Pi and OpenCode. Liquid also disclosed reward hacking during training (searching git history, detecting the sandbox) and says all results come from sealed environments.
Vendor-reported results vs Qwen3.6-35B-A3B: AutomationBench 27.69 vs 8.72, LiveClaw 48.09 vs 38.19, ToolSandbox 80.71 vs 75.02, BFCLv4 65.01 vs 65.00, IFEval 96.75 vs 93.06. It trails on SWE-Bench Verified (OpenCode) 51.00 vs 66.20, GPQA Diamond 74.42 vs 83.69 and WebArena Verified 26.75 vs 37.62, so Liquid recommends pairing it with retrieval. Decode speed is 184 tok/s on an Apple M5 Max and 114 tok/s on a Ryzen AI Max+ 395; a single B300 with SGLang exceeds 21K output tok/s.
Weights (plus GGUF, MLX, FP8 and ONNX builds) are on Hugging Face under the LFM Open License v1.0; Transformers >= 5.0.0 is required. Serve it via any OpenAI-compatible backend and follow Liquid's agent-harness guide for Hermes, OpenClaw or Pi. API pricing is $0.10 input ($0.05 cached) and $1.00 output per million tokens, and the model is also listed on OpenRouter.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.