Liquid AI releases LFM2.5-2.6B, a 2.69-billion-parameter local agent

Liquid AI released the weights of LFM2.5-2.6B on August 4, a text model designed less as an encyclopedia or general programmer than as an engine for small agents. It can reason, emit function calls, and work through multiple steps inside harnesses such as OpenClaw or Hermes Agent. Its size allows a quantized version to run on a phone or laptop without sending every token to an API.

The novelty is not a new architecture from scratch. LFM2-2.6B already combined 22 short-convolution blocks with eight attention blocks and already supported tools. The new checkpoint retains that 30-layer structure, but raises the reported pretraining budget from 10 trillion to about 34 trillion tokens, doubles the vocabulary to 128,000 units, extends context from 32K to 128K, and substantially changes post-training. Liquid has published both the base and post-trained models, along with Transformers, GGUF, MLX, and ONNX formats.

That turns a product claim into something developers and researchers can test themselves. It also exposes a distinction lost in the slogan “agents everywhere”: moving the model onto the device reduces one cloud dependency, but it does not remove the surrounding system.

Training inside the harness changes the skill being learned

The LFM2 technical report describes the hybrid architecture and training of the previous generation. For LFM2.5-2.6B, the main documentation so far is Liquid AI's release post and model card. They describe four stages: two rounds of supervised fine-tuning; specialization of several teacher models; on-policy distillation of those specialists into one student; and reinforcement learning inside agent environments.

The final stage is the most interesting. The model does not learn only to produce an isolated function call. Liquid runs it inside real harnesses, with each task in its own sandbox, so it can observe results, revise a plan, and finish research, writing, data-analysis, document, or coding work. Its reward combines a rubric graded by another model, programmatic checks, and a safety gate whose operation is not disclosed.

Training in the protocol used for deployment can reduce mundane formatting and coordination failures. It also narrows what the result establishes: an improvement inside OpenClaw or Hermes may reflect transferable skill, familiarity with their system messages, or some mixture of the two. Liquid has not published the training tasks, success distribution, a contamination analysis, or a checkpoint-specific safety card. The LFM2.5-2.6B model card itself advises against using it for agentic coding and knowledge-heavy work, two significant boundaries for a model presented in the announcement as a general agent.

The benchmarks support a narrower claim. In the provider's runs, the model beats larger Gemma variants and approaches Qwen3.5 on several instruction-following, tool-use, and agentic tests; the larger families retain an edge in coding and some knowledge tasks. BFCL v4 tests function selection and composition, multi-step interaction, memory, and search; ToolSandbox adds state and dependencies between tools. Those are relevant tests of whether a model speaks an agent's language. They do not measure whether its permissions are well designed, whether it resists malicious instructions returned by a website, or whether it can operate for a week without accumulating errors.

The numbers remain Liquid's results, not an independent reproduction. The post does publish temperatures and generation limits, and acknowledges using a different setting for Qwen because greedy decoding caused looping. On the official BFCL leaderboard, checked when this entry was published, LFM2.5-2.6B is not yet listed. Available weights make it possible to close that gap; they do not close it by themselves.

A 128K context window does not fit into one memory figure

Liquid reports 220 tokens per second on an M5 Max, 113 on a Ryzen AI Max+ 395 CPU, and 30 on a phone while using less than 2.5 GB of memory. Those are attractive measurements, but the sentence combines different choices. The official GGUF files take about 1.59–1.67 GB in Q4, 1.94 GB in Q5, and 2.22 GB in Q6; the BF16 model takes 5.40 GB before reserving memory for context, the harness, or tools. The 2.5 GB threshold therefore necessarily describes a quantized variant under a particular load, not every possible use of the model.

The cache retaining context also grows with the conversation. The page provides a precise methodology for its H100 test—BF16, 1,024 input tokens, up to 256 output tokens, and three runs—but not the same written detail for the CPU and phone chart, nor a quality, latency, and memory curve out to 128K. A window accepted by the configuration does not prove that the model reliably retrieves every fact spread through it, or that a device maintains its speed as the history approaches the maximum.

The useful project test would not be “does it start on my phone?” but “does it finish my workflow with this quantization, this history, and this tool set?” It should measure end-to-end success, energy use, memory growth, quantization loss, and recovery after a failed call. For Spanish and the other fifteen declared languages, it is also worth testing representative instructions and tool arguments: the announcement does not break results down by language.

Local does not mean isolated, free, or unrestricted

A local agent can keep its prompt, documents, and reasoning on the device. That is a real gain for offline tasks or data that should not reach a model provider. But if the agent searches the web, queries a CRM, or sends an email, the arguments to those tools leave the device. Privacy depends on the whole path: permissions, logs, remote services, and the data each tool returns to context.

Cost does not disappear either. What goes away is an API's per-token price, not energy, hardware, battery use, maintenance, or the time spent reviewing actions. Precisely because a small model allows many loops to run in parallel, a cheap mistake can be repeated at scale. The training sandbox does not travel inside the weights; whoever deploys the agent must still constrain its network, files, credentials, spending, and irreversible actions.

There is also a legal restriction obscured by the promotional language. Liquid calls the weights “open” and says they can be deployed “without restrictions.” Its LFM Open License 1.0, however, does not license commercial use by an entity with annual revenue of $10 million or more. The files can be downloaded, modified, and inspected, but they do not carry a permissive license for every use. “Open weights” describes technical access here; it is not synonymous with open-source software or universal commercial freedom.

My reading is that LFM2.5-2.6B matters because of where it places sufficient capability, not because it replaces much larger models. A small engine that selects tools, maintains a stateful conversation, and works offline can handle private, repetitive, bounded automation where a remote API would be costly, slow, or unacceptable. Its reasonable role resembles a specialized operator more than an authority: execute within verifiable limits, ask for help outside its domain, and leave a record of its actions.

The weights make it possible to find out whether that proposition survives outside Liquid's charts. The decisive signal will be third parties reproducing performance on specific devices and publishing completion and failure rates under quantization, long contexts, other languages, and hostile inputs. Until then, the release materially expands what can be tested locally; it does not turn a device into a free cloud or make an agent safe by default.

Sources

• Liquid AI, LFM2.5-2.6B: Deploy Agents Everywhere, August 4, 2026.
• Liquid AI, LFM2.5-2.6B model card, configuration, and weights and GGUF variants, accessed August 9, 2026.
• Liquid AI, LFM2 Technical Report, version accessed August 9, 2026.
• Liquid AI, LFM Open License 1.0, accessed August 9, 2026.
• UC Berkeley, Berkeley Function-Calling Leaderboard v4, and Lu et al., ToolSandbox, accessed August 9, 2026.