- Published on
An On-Device Agentic Model With 128K Context, Tool Calling, And Open Weights
- Authors
- Name
- DeepHorus
Liquid AI has released LFM2.5-2.6B, a compact agentic model designed to run entirely on-device, with open weights and a 128K context window. The model can plan, call tools, and work through multi-step tasks on phones, laptops, PCs, and robots, according to the company. With 2.69 billion parameters and a 131,072-token context, it aims to bring powerful agentic capabilities to hardware users already own.
The release comes as demand grows for edge AI that balances capability with privacy and cost. Liquid AI frames LFM2.5-2.6B as a direct response to the need for local inference, where data never leaves the device and the marginal cost of each run is near zero. The company previously shipped LFM2.5-8B-A1B, and this smaller model extends the same design philosophy to more constrained environments.
Two checkpoints are now public on Hugging Face under the lfm1.0 license: LFM2.5-2.6B-Base for fine-tuning, and LFM2.5-2.6B post-trained for agentic workloads. Both ship in native, GGUF, MLX, and ONNX formats, with day-one support in llama.cpp, vLLM, SGLang, and LM Studio. The model covers 16 languages and is text-only, pre-trained on roughly 34 trillion tokens.
Architecturally, LFM2.5-2.6B uses 22 double-gated short convolution blocks plus 8 grouped-query attention blocks across 30 layers. Liquid AI doubled the vocabulary to 128,000 by extending its existing tokenizer in place, then ran a dedicated mid-training phase to extend context to 128K tokens. The post-training recipe involves four stages: two supervised fine-tuning rounds, teacher specialization per domain, multi-domain on-policy distillation, and agentic reinforcement learning with GRPO inside real harnesses like Hermes Agent and OpenClaw.
The practical impact is significant for developers and enterprises. A single NVIDIA H100 SXM5 can serve roughly 1.3 billion tokens per day, while the model decodes at 220 tokens per second on an M5 Max using under 2.5 GB of memory. This allows solo developers to pilot agentic workloads on existing hardware, and mid-market teams to self-host on one GPU. Regulated industries such as healthcare, financial services, and defense benefit most because no prompt reaches a third-party API.
Analysts have noted early benchmarks where LFM2.5-2.6B leads every instruction-following metric reported, including ToolSandbox, Multi-IF, and IFStruct, compared with gemma-4-E4B-it and Qwen3.5-9B. Tool-use scores are competitive with models nearly four times its size, though larger models still outperform it on coding, with LiveCodeBenchv6 at 59.41 versus 69.86 for Qwen3.5-9B. The open-weights stance has drawn interest from the ML community, with 150,000-plus members across Liquid AI's channels.
Looking ahead, Liquid AI expects OEMs to push the same weights to device fleets via GGUF and ONNX, enabling on-device assistants, offline document triage, and continuous background agents. Fine-tuning support via LoRA with TRL and Unsloth should further accelerate adoption. The release underscores a broader industry shift toward on-device agentic AI, where models are small enough to run locally yet capable enough to handle complex, multi-step tasks without cloud dependency.
