Liquid, an AI startup founded in 2023 by former MIT computer scientists, has released LFM2.5-2.6B, a 2.6-billion-parameter open-weight language model designed for agentic workloads and local deployment. According to Liquid’s release materials and a follow-up interview with VentureBeat, the model can run entirely on local hardware — from smartphones and laptops down to a Raspberry Pi — without relying on cloud inference or GPUs. That capability targets edge AI use cases and organisations that need low-latency or privacy-preserving options.
Target use cases
Liquid positions LFM2.5-2.6B for high-volume, well-defined agentic tasks that run locally: tool calling, document management, calendar and workflow automation, and always-on background routines. The model is also intended for connectivity-limited environments such as vehicles and robotics. Liquid notes that coding-heavy work is still better handled by larger models.
Both the post-trained model and a base checkpoint for fine-tuning (LFM2.5-2.6B-Base) are available on Hugging Face. Day-one support is provided for major inference stacks including llama.cpp, MLX, vLLM, SGLang and ONNX, making the model deployable across consumer hardware, enterprise infrastructure and embedded systems. Liquid also provides an open-source fine-tuning framework called LEAP.
Technical specs and performance
LFM2.5-2.6B has 2.6 billion parameters, supports a 128,000-token context window, and includes native tool calling. The model name reflects its generation (2.5) and parameter count (2.6B).
Company-reported performance figures indicate decoding throughput of roughly 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 while using under 2.5 GB of memory. On smartphones the company reports around 30 tokens per second. Apollo, Liquid AI’s mobile app, can be used to try the model on phones.
On the high end, Liquid reports nearly 15,000 output tokens per second on a single Nvidia H100 GPU under sustained concurrent load — approximately 1.3 billion tokens per day on one card. These are vendor benchmarks and have not been independently verified.
Liquid intentionally designed the underlying LFM2 architecture around real-world CPU performance rather than GPU benchmarks. Maxime Labonne, Liquid AI’s head of post-training, said the model runs “very, very well” on CPUs and cited Raspberry Pi demos that show fast performance on that platform.
Trained for agents rather than chat
Liquid developed LFM2.5-2.6B with the assumption that language models increasingly run inside agent frameworks rather than traditional chat interfaces. The model was pretrained on roughly 34 trillion tokens, the vocabulary was doubled to 128K to better support non-Latin scripts, and a mid-training phase extended the context window to 128K tokens for long-running agent workflows.
Post-training followed a four-stage pipeline: supervised fine-tuning; teacher specialization (training separate expert models for domains such as instruction following, math, code, and tool use); multi-domain on-policy distillation (MOPD) to merge those experts into a single student model; and agentic reinforcement learning. During the final stage the model was trained inside production agent harnesses — including Hermes Agent and OpenClaw — on realistic productivity tasks that exposed it to actual tools, system prompts and interaction patterns.
Labonne described the pipeline changes as a “happy accident” that yielded broader improvements, including gains in math, instruction following and code performance.
Co-designing the harness and model
Liquid also built its own agent harness rather than relying only on existing frameworks, and demonstrated the model running inside that harness on a phone, planning and calling tools entirely on-device. The company cited two reasons: a lack of phone-native harnesses and the desire for a different interaction model. Liquid wants proactive, background agents that monitor context (calendar, user activity) and act autonomously.
Co-design lets the harness compensate for model weaknesses — the harness provides assistance to make the system more reliable. Still, the model works out of the box with established harnesses such as Hermes Agent, OpenClaw and Pi, and can be served behind any OpenAI-compatible endpoint.
Deployment economics and fine-tuning
Labonne argued the release shifts what small models can do for enterprises. Previously, local models were mainly narrow, fine-tuned specialists; agentic capability changes that calculus because the same model can be reused for multiple tasks by swapping the harness and tools rather than the model itself. He recommends fine-tuning for production deployments when possible: a well-fine-tuned LFM2.5-2.6B can match GPT- and Claude-class performance for many tasks, he said, and the barrier to entry for fine-tuning is now low.
Benchmarks and comparisons
Liquid published comparisons against models that enterprises might consider for edge deployments: Google’s Gemma 4 E2B (5.1B) and E4B (8B), and Alibaba’s Qwen3.5-4B (4.7B) and Qwen3.5-9B (9.7B). A separate test by the local AI client platform Atomic Chat found that LFM2.5-2.6B completed 35 tool calls to finish three tasks (checking weather and local time in six cities, converting a budget into six currencies, and checking/booking hotels for a date) 3.7 times faster than DeepSeek-V4-Flash (284B parameters).
LFM2.5-2.6B is text-only, dense and specialized for agentic work; Liquid ships separate vision and audio variants for the LFM family. Gemma and Qwen models tend to be multimodal and pursue different training trade-offs: Gemma uses Per-Layer Embeddings and reports “effective” sizes, while Qwen’s larger models emphasize reinforcement learning for frontier-style reasoning.
According to Liquid’s published numbers, the LFM2.5-2.6B model leads instruction-following benchmarks (IFBench, Multi-IF, IFStruct) and nearly every tool-use benchmark — e.g., 77.83 on ToolSandbox versus 76.44 for Qwen3.5-9B — and performs strongly on agentic evaluations and knowledge benchmarks that penalize hallucinations. Qwen models retain an edge on math (AIME25) and coding (LiveCodeBench), where larger parameter counts still help.
Licensing and commercial terms
LFM2.5-2.6B is distributed under the LFM Open License v1.0, which allows use, modification and redistribution — including commercial use — for organisations with under $10 million in annual revenue. Commercial use by larger companies requires a separate arrangement with Liquid AI; qualified nonprofits are exempt from the threshold for non-commercial and research purposes.
Labonne described the licensing as a pragmatic way to fund ongoing model development: “the models are really the moats, so we need to be sensible in the way that we license them; otherwise, we cannot make money, so we can't make more models.” He acknowledged limits in enforcement: if a large enterprise quietly deploys the weights, the onus is on legal teams and the company requests that such organisations contact Liquid.
Liquid pairs the licensed release with freely published research, including structured-output evaluations and a training technique aimed at reducing repetition loops common in small models.
Partnerships and market implications
The release coincided with a strategic partnership announced with MacPaw, the Ukrainian company behind CleanMyMac and Setapp. Liquid will design and fine-tune foundation models for Eney, MacPaw’s macOS assistant, to run locally on Apple silicon through MacPaw’s Elix inference engine and Mnemos memory layer; results are expected later this year.
Labonne pointed to the MacPaw deal as validation of the small-model argument: MacPaw chose Liquid in part because the model is small and fits their memory budget.
The announcement arrives as hardware vendors, OS developers and enterprise software companies increasingly invest in local AI execution and as agent harnesses proliferate. Liquid’s bet is that deployment economics — latency, privacy, flexibility and marginal inference cost — rather than raw scale, will define an important enterprise segment: continuously running agents everywhere at essentially zero marginal token cost.
Ultimately, whether small, optimized agent models capture a significant share of enterprise AI will depend on operational reliability more than benchmark rankings. Liquid AI’s release suggests the next frontier may be building models that are compact and capable enough to run where enterprise workflows already live.
