Skip to content
NewsModel release

Liquid AI ships LFM2.5-2.6B, a 2.6B open-weight on-device agent model with 128k context

· by Pondero Newsdesk

The short version

Liquid AI released LFM2.5-2.6B on August 4, an open-weight 2.6-billion-parameter model that handles 128k-token contexts, calls tools natively, and runs on consumer hardware including phones, with no per-token billing.

Liquid AI ships LFM2.5-2.6B, a 2.6B open-weight on-device agent model with 128k context

Every token on a cloud AI API carries a charge. Liquid AI published LFM2.5-2.6B on August 4 to remove that as a variable: an open-weight 2.6-billion-parameter model that runs entirely on-device, handles 128k-token contexts natively, and calls tools without a cloud API. Per the company's published benchmark table, it leads instruction-following evaluations against models up to 9.7 billion parameters.

What

LFM2.5-2.6B has 2.6 billion parameters and was pre-trained on roughly 34 trillion tokens, per Liquid AI's announcement. The 128k context window came from a dedicated mid-training extension phase. Liquid AI also doubled the model's vocabulary to 128,000 tokens by extending the existing tokenizer in place rather than retraining from scratch.

Post-training ran through four stages. Two supervised fine-tuning rounds started the sequence, with the second weighted toward tool use, agentic traces, and web search tasks. Teacher Specialization followed: domain experts trained via reinforcement learning from the same shared SFT checkpoint. The third stage, Multi-Domain On-Policy Distillation (MOPD), let the student roll out under its own policy while a matching specialist teacher provided token-level feedback per prompt. The fourth stage, Agentic RL, optimized behavior inside real agent harnesses including Hermes Agent and OpenClaw.

Per Liquid AI's inference benchmarks, the model decodes at 220 tokens per second on an Apple M5 Max, 113 tokens per second on an AMD Ryzen AI Max+ 395, and 30 tokens per second on a phone, while staying under 2.5 GB of memory. On a single NVIDIA H100 SXM5, it reached roughly 15,000 output tokens per second at high concurrency, approximately 1.3 billion tokens per day. Inference runs on llama.cpp, MLX, vLLM, SGLang, and ONNX from day one. Both the base (LFM2.5-2.6B-Base) and post-trained checkpoints are available on Hugging Face.

Against Gemma-4-E2B-it (5.1B), Gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B), and Qwen3.5-9B (9.7B), LFM2.5-2.6B led every instruction-following evaluation in the published table: IFBench 59.17 versus 56.47 for Qwen3.5-9B; Multi-IF 80.07 versus 62.55; IFStruct 85.49 versus 78.50. On ToolSandbox it led (77.83 versus 76.44 for Qwen3.5-9B). On BFCLv4 it trailed Qwen3.5-9B by 3.25 points. Coding benchmarks remained the clearest gap favoring the larger models.

Why it matters

Per-token billing shapes which agentic architectures are economically practical. Continuous background agents and round-the-clock automated research look different at zero marginal token cost than at scale-with-usage pricing. Liquid AI's blog states that removing the per-token cost "changes how developers build," per the announcement.

The benchmark figures matter on their own terms. LFM2.5-2.6B scores 80.07 on Multi-IF against 62.55 for Qwen3.5-9B, a model with roughly 3.7x more parameters. For teams evaluating small models for edge agentic workloads, the data shows on-device can match or beat hosted models that are substantially larger on the capabilities that drive agentic task completion.

On-device inference satisfies data residency requirements that hosted APIs cannot. Prompts and completions stay on the device, which matters for healthcare, legal, and financial services deployments where third-party data transmission is restricted or prohibited.

What to watch next

Community benchmarks against Llama 3.2-3B and Phi-3.5-mini will show where LFM2.5-2.6B stands relative to the most widely deployed open models in its size class. Submission to the LMSYS Chatbot Arena would provide an independent quality signal on conversational performance outside the company's own evaluation table. Published latency data from Raspberry Pi deployments would test whether the 30-tokens-per-second phone figure holds on lower-power hardware.

Sources