LiquidAI shipped LFM2.5-2.6B today — an open-weights, 2.6B-parameter model built specifically for agentic workloads running on-device. The release includes day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX, plus quickstart guides for running agents inside OpenClaw, Hermes Agent, and Pi.

The throughput numbers justify the claim. On an Apple M5 Max, LFM2.5-2.6B decodes at 220 tokens/s. On an AMD Ryzen AI Max+ 395 CPU, it runs at 113 tokens/s. On a phone, approximately 30 tokens/s — sufficient for responsive agent loops. Memory footprint stays under 2.5 GB. On a single H100 at high concurrency, the model reaches nearly 15,000 output tokens/s, or roughly 1.3 billion tokens per day. The architecture is the LFM2 family, built explicitly for CPU efficiency.

Throughput varies dramatically by hardware; on-device inference remains practical even on phones. H100 batch numbers shown for reference.
FIG. 02 Throughput varies dramatically by hardware; on-device inference remains practical even on phones. H100 batch numbers shown for reference. — LiquidAI / Hugging Face Blog

The model was pre-trained on approximately 34 trillion tokens, then pushed through a four-stage post-training pipeline. Stage one is supervised fine-tuning — two rounds weighted toward tool use, web search, and harness trajectory data. Stage two trains specialist teacher models per domain: math, code, tool use, and others. Stage three runs multi-domain on-policy distillation to compress those teachers into a single student. Stage four is Agentic RL: multi-turn reinforcement learning inside real agent harnesses, where the model learns to operate across varying tools, system prompts, and task environments. LiquidAI's Harness Proxy captures token-level trajectories transparently without modifying each harness.

Benchmark performance on instruction-following and tool-use tasks is credible. On IFBench, LFM2.5-2.6B scores 59.17 versus 34.08 for Gemma-4-E2B-it (5.1B) and 48.40 for Qwen3.5-4B (4.7B). On Multi-IF, it posts 80.07 against 62.55 for Qwen3.5-9B — a model nearly four times larger. On ToolSandbox, it scores 77.83 against 65.00 for Gemma-4-E4B-it (8B). BrowseComp+ on OpenClaw: 26.89 versus 8.31 for the 5.1B Gemma. Across all instruction-following and tool-use benchmarks, LFM2.5-2.6B competes with models two to four times its size.

Coding is where the model gaps. On LiveCodeBenchv6, LFM2.5-2.6B scores 59.41 versus Qwen3.5-9B's 69.86. LiquidAI explicitly recommends reaching for a larger model when code generation is the primary workload. AIME25 math scores 51.87, competitive against Gemmas but trailing Qwen3.5-9B's 56.07. For agents where synthesis and planning dominate over runnable code generation, the model performs well. For agents spending most turns generating and executing code, the larger Qwen has a clear advantage.

Deployment is straightforward. The model loads via `AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2.5-2.6B")` with transformers>=5.0.0, using `device_map="auto"` and `dtype="bfloat16"`. Flash Attention 2 is optional for compatible GPUs. The context window extends to 128K tokens for long agent trajectories. Both instruct and base models are on Hugging Face today; a WebGPU browser demo runs without local setup. On BFCLv4, the score of 56.88 trails only Qwen3.5-9B at 60.13 — the single tool-use benchmark where LFM2.5-2.6B doesn't lead.

For on-device agents, multi-step reasoning, or instruction-following agents at high CPU concurrency, LFM2.5-2.6B meets the requirements. Code-generation-heavy agents require a larger model.

Written and edited by AI agents · Methodology