|

Liquid AI Releases LFM2.5-2.6B: An On-Device Agentic Model With 128K Context, Tool Calling, And Open Weights

ByRicardo

Liquid AI launched LFM2.5-2.6B, an agentic mannequin that runs completely on-device. It plans, calls instruments, and works by multi-step duties on telephones, laptops, PCs, and robots. The mannequin has 2.69B complete parameters, a 131,072-token context window, and a 128,000-token vocabulary. Pre-training used roughly 34 trillion tokens. Two checkpoints shipped: LFM2.5-2.6B-Base for fine-tuning, and LFM2.5-2.6B post-trained for agentic workloads. Because inference stays native, information by no means leaves the gadget and the marginal value of every run is close to zero. Liquid AI reviews tool-use and instruction-following scores aggressive with fashions almost 4 instances its dimension.

Is it deployable

The reply is Yes. Both checkpoints are public on Hugging Face underneath the lfm1.0 license. Weights ship in native, GGUF, MLX, and ONNX codecs, with day-one assist in llama.cpp, vLLM, SGLang, and LM Studio.

  • Which corporations: Solo builders and startups can pilot on {hardware} they already personal. The mannequin decodes at 220 tokens/s on an M5 Max in underneath 2.5 GB. Mid-market groups can self-host on one GPU: a single NVIDIA H100 SXM5 serves roughly 1.3B tokens per day. Enterprises and OEMs can push the identical weights to gadget fleets by GGUF and ONNX. Fine-tuning is obtainable by way of LoRA with TRL and Unsloth.
  • Which industries: Liquid AI targets automotive, consumer electronics, industrial robotics, healthcare, financial services, e-commerce, and defense. Regulated and air-gapped settings profit most, since no immediate reaches a third-party API.
  • Applications: Liquid AI recommends agentic workloads, instrument use, information extraction, RAG, and long-context workflows. Practical builds embody on-device assistants, offline doc triage over 128K inputs, type and bill extraction, robotics command parsing, and background brokers that run repeatedly with out per-token value. Liquid AI explicitly doesn’t advocate the mannequin for agentic coding or knowledge-heavy duties.

Architecture and coaching finances

LFM2.5-2.6B has 2.69B complete parameters throughout 30 layers. The stack is 22 double-gated brief convolution blocks plus 8 grouped-query consideration blocks. Vocabulary dimension is 128,000 and context size is 131,072 tokens. Pre-training used roughly 34 trillion tokens.

Liquid AI doubled the vocabulary to 128K by extending the existing tokenizer in place fairly than retraining from scratch. A devoted mid-training part extends context to 128K. The mannequin covers 16 languages and is text-only.

Four-stage post-training

The base checkpoint turns into an agent by 4 levels.

  • First, two consecutive supervised fine-tuning rounds, with an SFT combine roughly seven instances the dimensions used for LFM2.5-8B-A1B.
  • Second, trainer specialization: one skilled per area, skilled with reinforcement studying with verifiable rewards.
  • Third, multi-domain on-policy distillation, the place the coed rolls out underneath its personal coverage and every immediate routes to its area trainer.
  • Fourth, agentic reinforcement studying with GRPO inside actual harnesses, together with Hermes Agent and OpenClaw.

Benchmarks

Liquid AI in contrast LFM2.5-2.6B in opposition to gemma-4-E2B-it (5.1B), gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B) and Qwen3.5-9B (9.7B).

Benchmark LFM2.5-2.6B gemma-4-E4B-it Qwen3.5-9B
ToolSandbox 77.83 65.00 76.44
Multi-IF 80.07 77.35 62.55
IFStruct 85.49 76.65 78.50
IFBench 59.17 39.24 56.47
BFCLv4 56.88 46.39 60.13

It leads each instruction-following benchmark reported and almost each instrument use benchmark, trailing Qwen3.5-9B solely on BFCLv4. Coding is the place bigger fashions maintain an edge: LiveCodeBenchv6 is 59.41 versus 69.86 for Qwen3.5-9B.

Interactive explainer