Liquid AI Releases LFM2.5-24B-A2B for Local Agents

Liquid AI released a 24B local-agent model on October 13

Liquid AI announced LFM2.5-24B-A2B on October 13, 2026 as the final and largest model in its LFM2.5 family. The post-trained model is available now as LiquidAI/LFM2.5-24B-A2B, alongside LiquidAI/LFM2.5-24B-A2B-Base. The public repository is ungated and contains ten Safetensors weight shards; Liquid AI also publishes official GGUF quantizations for llama.cpp.

This is an artifact release, not a newly priced managed API. Builders download and serve the weights with Transformers, llama.cpp, MLX, vLLM, SGLang or ONNX. When reviewed, Hugging Face did not list a hosted inference provider for the official GGUF repository. Hardware, serving, monitoring and incident costs therefore remain with the operator.

Only 2.45B parameters are active, but the full model still occupies memory

The model card identifies 24.0 billion total and 2.45 billion active parameters. Its 40-layer architecture combines 30 double-gated short-convolution blocks with 10 grouped-query-attention blocks. Each mixture-of-experts block has 64 experts and routes every token through four. The context window is 131,072 tokens, up from 32K in its predecessor, and the vocabulary doubles to 128,000 tokens.

The active count describes computation per token; it does not turn the model into a 2.45B download. Hugging Face reports about 23.97B BF16 parameters, and the official GGUF page displays 13.6 GB for Q4_0, 14.5 GB for Q4_K_M, 17 GB for Q5_K_M, 19.7 GB for Q6_K, 25.5 GB for Q8_0 and 48 GB for BF16 or F16. Long prompts also consume KV-cache memory. Size hardware from the actual quantization, context, concurrency and runtime—not the active-parameter headline.

The agent benchmark results are promising and provider-run

Liquid AI evaluated the model with vLLM on 16 benchmarks. It reports 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, 80.71 on ToolSandbox versus 75.02, and 96.75 on IFEval versus 93.06. It also reports 27.69 on AutomationBench, more than three times the next score shown in its table. These figures support a test hypothesis for tool use; they do not establish independent superiority on a production workload.

The same table shows where the model does not lead: 26.75 on WebArena Verified versus Qwen’s 37.62, 74.42 on GPQA Diamond versus 83.69, and 51.0 on SWE-bench Verified with OpenCode versus 66.2. AccessAllGPT did not reproduce any score or inspect all prompts, trajectories, failures and scoring decisions. Compare pinned harnesses at equal budgets, tool permissions, timeouts and retry rules before making a model decision.

The speed claims need workload-specific reproduction

Liquid AI reports 184 output tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, calling the model up to 1.7 times faster than the compared mixtures of experts on-device. It also reports more than 21,000 aggregate output tokens per second at high concurrency on one NVIDIA B300. These are vendor measurements, not AccessAllGPT tests.

The headline numbers answer different questions: single-device interactive decode and high-concurrency accelerator throughput. They do not provide a universal latency result. A useful local test records exact hardware, quantization, runtime commit, prompt and output lengths, context occupancy, batch and concurrency, time to first token, inter-token latency, p50/p95/p99 completion time, memory, energy, tool-call accuracy and accepted-task cost.

Reasoning and tool calls do not create an authorization boundary

Liquid AI says the post-training recipe included supervised fine-tuning, specialist teachers, two on-policy distillation stages and reinforcement learning inside agent harnesses including OpenClaw, Hermes Agent, Pi and OpenCode. Its documented chat template supports tools and emits an explicit thinking phase before the final answer.

That training does not make generated tool calls trustworthy. Keep credentials, scopes, approvals, argument validation, network policy, rate limits and irreversible-action checks in host code. Treat model reasoning as untrusted data, redact it where retention would expose sensitive inputs, and test malformed tool results, prompt injection, repeated actions and recovery after partial failure.

The license excludes commercial use by larger legal entities

The repository uses LFM Open License v1.0. It grants copyright and patent permissions subject to attribution and redistribution conditions, but states that commercial use by a legal entity with annual revenue of $10 million or more is not licensed. Its legal-entity definition includes organizations under common control. Liquid AI’s models page describes download and commercial use as free only until a company passes that threshold.

Open weights therefore describes artifact access, not unrestricted production rights. Confirm the applicable entity, revenue, use and distribution path with counsel. Organizations at or above the threshold should obtain a separate license or choose an alternative before commercial deployment; a successful technical evaluation does not cure a rights gap.

What local-agent teams should do next

Trial LFM2.5-24B-A2B when local data control, offline operation or predictable hardware ownership matters and the model fits both the license and memory budget. Freeze the repository commit, runtime and quantization; scan dependencies and weight provenance; then compare it with the smallest credible baseline on real tool traces, long-context cases and adversarial instructions.

Constrain any pilot to low-impact tools with explicit approval and rollback. Wait when commercial rights, hardware fit, benchmark reproducibility or long-context behavior is unresolved. Reject it when the model cannot meet task quality, latency, memory, energy, security or total-cost gates—or when a generated tool call would be allowed to authorize itself.

Disclosures

AccessAllGPT received no model files, hardware, account, credits, benchmark data, briefing, review, payment or compensation for this article. Liquid AI, Hugging Face, Apple, AMD, NVIDIA and the named runtime and model projects did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank LFM2.5-24B-A2B. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Liquid AI or organizations cited. Publication-wide relationships are listed on the disclosures page.