Log in Sign up
Business Chemistry-Native AI Startup Albert Raises $55 Million Series B Led by TCV Chatbots Liquid AI Ships Downloadable LFM2.5-24B-A2B, a Reasoning Model That Activates Just 2 Billion Parameters Per Token Coding Assistants Anthropic Replaced Git With Its Own System After Daily Code Merges Hit 30,000 Gemini Google Opens Gemini 4 Argon to Paying Users Worldwide, With $2 Intro Pricing That Doubles After November 30 News Ubitium Raises $14 Million to Turn Working Silicon Into a Universal Processor for Physical AI AI Agents American Express Wants to Check What You Meant Before an AI Agent Buys It Apple Intelligence Apple Puts a 7-Inch Siri AI Hub on Your Wall: HomeView Costs $299 and Ships November 13 Business Samsung and SK Hynix Slide as Chinese Memory Maker CXMT Signals a Denser 4F-Squared DRAM Design Business Chemistry-Native AI Startup Albert Raises $55 Million Series B Led by TCV Chatbots Liquid AI Ships Downloadable LFM2.5-24B-A2B, a Reasoning Model That Activates Just 2 Billion Parameters Per Token Coding Assistants Anthropic Replaced Git With Its Own System After Daily Code Merges Hit 30,000 Gemini Google Opens Gemini 4 Argon to Paying Users Worldwide, With $2 Intro Pricing That Doubles After November 30 News Ubitium Raises $14 Million to Turn Working Silicon Into a Universal Processor for Physical AI AI Agents American Express Wants to Check What You Meant Before an AI Agent Buys It Apple Intelligence Apple Puts a 7-Inch Siri AI Hub on Your Wall: HomeView Costs $299 and Ships November 13 Business Samsung and SK Hynix Slide as Chinese Memory Maker CXMT Signals a Denser 4F-Squared DRAM Design

Liquid AI Ships Downloadable LFM2.5-24B-A2B, a Reasoning Model That Activates Just 2 Billion Parameters Per Token

Chatbots · By: AIReport Admin · 1 hour ago · 👁 3
Liquid AI Ships Downloadable LFM2.5-24B-A2B, a Reasoning Model That Activates Just 2 Billion Parameters Per Token

Boston startup Liquid AI released LFM2.5-24B-A2B on Monday, a mixture-of-experts reasoning model with downloadable weights on Hugging Face that activates only 2 billion of its 24 billion parameters per token. The context window grows to 131K tokens and pre-training doubled to roughly 33 trillion tokens. Benchmarks are company-reported: it leads on tool use but trails Qwen on coding and web tasks.

Liquid AI, the Boston-based startup founded by former MIT CSAIL researchers, released a new open-weights model on Monday aimed at developers who want agentic reasoning running on their own hardware rather than behind a hosted API. LFM2.5-24B-A2B, together with its Base variant, is available for download on Hugging Face.

The model is a mixture-of-experts design that activates just 2 billion of its 24 billion parameters per token, keeping its predecessor's 40-layer architecture with 64 experts per block and four selected at a time. Unlike LFM2-24B-A2B, the new release is reasoning-only: it generates an explicit reasoning trace before its final answer by design rather than as an optional mode.

The specification changes are substantial. The context window expands from 32K to 131K tokens, and the tokenizer vocabulary doubles from 65K to 128K, which Liquid says improves representation of non-Latin writing systems. Pre-training grew from roughly 17 trillion to 33 trillion tokens, with later phases emphasizing reasoning, math, code and instruction following, and the final reinforcement-learning stage trained the model inside agent frameworks including OpenClaw, Hermes Agent, Pi and OpenCode.

On Liquid's own evaluations — run across 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, all through vLLM, and all company-reported rather than independently verified — the model led or matched the comparison group on five of eight tool-use and agent benchmarks and led on all three instruction-following tests. On LiveClaw it scored 48.09 against Qwen's 38.19.

The honest caveat is that the targeted strength is not a broad win. Qwen retained a substantial advantage elsewhere: on WebArena Verified, Liquid scored 26.75 versus Qwen's 37.62, and on SWE-Bench Verified evaluated through OpenCode the scores were 51.00 and 66.20. Liquid acknowledges that larger models keep an edge on knowledge-heavy and coding tasks and recommends pairing LFM2.5 with retrieval or web search for knowledge-intensive work — positioning it as a tool-using component, not a universal replacement.

That positioning is the interesting part of the release. The on-device agent market is filling in fast: Apple, Google and Qualcomm are all pushing small-model inference onto consumer hardware, while startups compete on efficiency per watt rather than raw benchmark peaks. A 24B-total, 2B-active model with a 131K context that can run on a single workstation GPU — or a high-end mini PC — is a plausible sweet spot for agents that call tools all day where API latency and per-token bills compound.

What to watch next is independent reproduction. Liquid's transparent disclosure of where its model loses — unusual among model releases — makes the claims easier to trust, but the numbers still come from the company's own harness. Third-party evaluations on the tool-use benchmarks, and real-world token throughput on consumer hardware, will determine whether LFM2.5 becomes a default for local agent stacks or another interesting checkpoint on the shelf.

Comments (0)

Log in to join the discussion

Log In

No comments yet