Liquid AI Releases LFM2.5-2.6B, a 2.6B-Parameter Agentic Model with 128K Context for On-Device Use

Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter model built specifically for on-device agentic tasks, according to the company's Hugging Face model card. The model extends the LFM2 hybrid architecture with a 128K-token context window and a multi-stage agentic post-training pipeline.

Key Specifications

• Total parameters: 2.69B
• Architecture: 30 layers (22 double-gated short convolution blocks + 8 grouped-query attention blocks)
• Training budget: ~34 trillion tokens
• Context length: 131,072 tokens
• Vocabulary size: 128,000 tokens
• Languages supported: 16, including English, Arabic, Chinese, French, German, Japanese, Korean, Russian, and Hindi
• Model type: Reasoning model — the model always generates a <think> block before answering, per Liquid AI's documentation

The release includes a base checkpoint (LFM2.5-2.6B-Base) for fine-tuning, plus the post-trained agentic version, and is distributed in GGUF, ONNX, and MLX formats for deployment across CPU, GPU, cloud, and mobile environments.

Training Pipeline

Liquid AI pre-trained the model on roughly 34 trillion tokens, followed by a mid-training phase that extended context to 128K tokens. Post-training involved four stages: two rounds of supervised fine-tuning, per-domain teacher specialization, multi-domain on-policy distillation, and what the company calls "agentic reinforcement learning" — training the model directly inside popular agentic harnesses so it learns their tools, system prompts, and interaction patterns.

Benchmark Claims

Liquid AI claims LFM2.5-2.6B outperforms several larger sub-10B models on agentic benchmarks. According to the company's published results, LFM2.5-2.6B scored 56.88 on BFCLv4 versus 60.13 for the 9.7B Qwen3.5-9B, and 85.49 on IFStruct versus 78.50 for the same larger model. On IFBench, LFM2.5-2.6B scored 59.17, ahead of Qwen3.5-4B's 48.40 and Gemma-4-E4B-it's 39.24. On AIME25, it scored 51.87, trailing Qwen3.5-9B's 56.07 but ahead of both Gemma variants tested (26.33 and 34.27). These figures come directly from Liquid AI and have not been independently verified.

Inference Performance

Liquid AI reports decode speeds of 220 tokens/second on an Apple M5 Max and 113 tokens/second on an AMD Ryzen AI Max+ 395 CPU, using under 2.5 GB of memory. The company also claims the model reaches close to 15,000 output tokens per second at high concurrency on GPU, positioning it as the fastest model tested in its size class. Pricing is not applicable, as the model weights are freely available on Hugging Face; pricing not yet disclosed for any hosted inference option.

The model supports Pythonic function calling by default, with tool definitions passed via JSON schema in the system prompt or through the tokenizer's chat template. It is compatible with Transformers (v5.0.0+), vLLM, SGLang, llama.cpp, MLX, and LM Studio.

What This Means

LFM2.5-2.6B targets a specific niche: agentic workloads that need to run locally on phones, laptops, or edge hardware rather than in the cloud. A 2.6B model claiming near-parity with 9B+ models on tool-use and instruction-following benchmarks — if the numbers hold up under independent testing — would be notable for developers building on-device AI assistants where cloud inference costs or latency are prohibitive. The explicit agentic RL training inside real harnesses is a more deliberate approach than typical benchmark-chasing, but Liquid AI's own comparisons should be treated as directional rather than definitive until third-party evaluations confirm them. The company itself notes the model is not recommended for coding agents or knowledge-heavy tasks, signaling this is a narrow-purpose tool rather than a general-purpose competitor to larger frontier models.