LFM2.5-2.6B¶

Development line: project:lfm2-5-2-6b · thread lfm2-5-2-6b-development
Last event: 2026-08-04 · 1 dated since 2026-08-04 · Researched: 2026-09-05 · confidence: medium

What it is¶

LFM2.5-2.6B is a post-trained text-only checkpoint for builders deploying local agents through an OpenAI-compatible endpoint.

• Plans multi-step tasks and calls tools.
• Supports 16 languages and a 131,072-token context.
• Has GGUF, ONNX, MLX, QAD 4-bit, and DSpark deployment paths.

Development line¶

• 2026-08-04 — Liquid AI published the LFM2.5-2.6B model line. On 2026-08-04, Liquid AI linked a blog entry and Hugging Face resources for LFM2.5-2.6B. The linked resources included the model page, quantized-model discovery, a WebGPU space, and a playground so practitioners can run the checkpoints.

What changed¶

• 2026-08-04 — Liquid AI released the pre-trained LFM2.5-2.6B-Base and the agentically post-trained LFM2.5-2.6B.
• 2026-08-19 — Liquid AI added a QAD-trained Q4_0 GGUF for LFM2.5-2.6B to recover low-bit accuracy on the same Q4_0 runtime path.
• 2026-08-20 — Liquid AI released the 327.7M-parameter LFM2.5-2.6B-DSpark draft model for speculative decoding with the original checkpoint.

How to use this¶

Evaluate LFM2.5-2.6B on its linked model page and public demo surfaces as of 2026-08-04 when considering a compact Liquid AI model.

1. Choose the post-trained checkpoint for an agent, or LFM2.5-2.6B-Base for fine-tuning. Apply the documented chat template when calling it through Transformers. — https://huggingface.co/LiquidAI/LFM2.5-2.6B
2. For local llama.cpp deployments, start with the Q4_K_M GGUF. Serve it with its Jinja template enabled and expose the local OpenAI-compatible endpoint. — https://docs.liquid.ai/examples/agent-harnesses
3. Point Hermes, OpenClaw, or Pi at that endpoint. Enable tool-use enforcement or the equivalent tool-call parser in the serving backend. — https://docs.liquid.ai/examples/agent-harnesses
4. Attach LiquidAI/LFM2.5-2.6B-DSpark as the speculative draft model on compatible SGLang builds to cut latency. Retain the original 2.6B model as the target. — https://huggingface.co/LiquidAI/LFM2.5-2.6B-DSpark
5. Use the maintained WebGPU Space for browser-side evaluation before installing a local runtime. — https://huggingface.co/spaces/LiquidAI/LFM2.5-2.6B-WebGPU

Best practices¶

• Start GGUF deployments at Q4_K_M (1.67 GB). Move to Q8_0 for tool-heavy work or BF16 for fidelity and benchmarking. Reduce the served context to about 32K when memory is constrained. — https://docs.liquid.ai/examples/agent-harnesses
• Keep the model chat template and tool-calling configuration enabled. For llama.cpp, start with temperature 0.1, top-k 50, and repeat penalty 1.1. — https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF
• Do not select it as the default for agentic coding or knowledge-heavy tasks. Use it where local tool use, extraction, RAG, or long-context workflow fits the model's stated scope. — https://huggingface.co/LiquidAI/LFM2.5-2.6B
• Check commercial eligibility before distribution. The LFM Open License v1.0 does not license commercial use by a legal entity at or above US$10 million annual revenue. — https://huggingface.co/LiquidAI/LFM2.5-2.6B/blob/main/LICENSE

Superseded by this¶

• 2026-08-19 — Generic guidance to use the older post-training-quantized Q4_0 checkpoint is superseded where the QAD Q4_0 checkpoint is available. Liquid AI reports that the 2.6B QAD build retains 96.6% of its BF16 average while using the same Q4_0 deployment path.

Still unknown¶

• We ran no independent reproducible deployment or benchmark here, so performance figures remain vendor-reported.
• Chinese-language search found third-party reports but no dated first-party Chinese Liquid AI documentation, so we claim no Chinese operating guidance.
• The WebGPU Space was running when read. Liquid Playground redirected to sign-in, so unauthenticated Playground access could not be confirmed.

Sources¶

Agent brief¶

• Subject: project:lfm2-5-2-6b, thread lfm2-5-2-6b-development, 1 dated events 2026-08-04 → 2026-08-04.
• Practical note: As of 2026-08-04, practitioners can evaluate LFM2.5-2.6B through its linked model page and public demo surfaces when considering a compact Liquid AI model.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.