What is the best local LLM for on-device agents in 2026?

What is LFM2.5-24B-A2B?

LFM2.5-24B-A2B is Liquid AI's open-weight mixture-of-experts reasoning model for local agent work, released on 13 October 2026. It holds 24.0B total parameters while activating only 2.45B per token, which is why it stays usable on hardware that could not host a dense 24B model at interactive speed.

The model card lists 40 layers (30 double-gated short convolution blocks plus 10 grouped-query attention blocks), 64 experts per MoE block with top-4 routing, a 131,072-token context window, a 128,000-token vocabulary and 16 languages. Pre-training ran on roughly 33 trillion tokens with another 2 trillion of mid-training and 300 billion long-context tokens, up from about 17 trillion on the previous LFM2-24B-A2B.

Two changes matter to a buyer. First, this is a reasoning-only model: it always emits a thinking block before answering, the chat template inserts the think tag automatically, and there is no fast non-thinking mode to switch to. Second, it is text-only, and Liquid AI's own guidance says it is not recommended for knowledge-heavy tasks on its own, where it should be paired with retrieval or web search.

The post-training recipe is unusual enough to note. Liquid AI ran agentic reinforcement learning inside named agent harnesses: OpenClaw, Hermes Agent, Pi and OpenCode. Tool-calling behaviour is therefore tuned to the shapes those harnesses send, including Pythonic function calls wrapped in tool-call tokens.

Is it the best local LLM, or just the newest?

The benchmark table on Liquid AI's launch page compares LFM2.5-24B-A2B against three models with more parameters, and the vendor's headline claim is that it is competitive with Qwen3.6-35B-A3B on five of eight agentic and tool-use benchmarks. The useful part is the two-way reading, because the losses are as specific as the wins.

Where LFM2.5 leads, per Liquid AI: AutomationBench 27.69 against 8.72 for Qwen3.6-35B-A3B, 5.46 for Gemma-4-26B-A4B-it and 2.61 for Nemotron-3.5-Lightning 30B-A3B; LiveClaw 48.09 against 38.19, 41.25 and 37.72; BFCLv4 65.01 against 65.00, 57.09 and 53.78; ToolSandbox 80.71 against 75.02, 72.09 and 75.13; and the banking slice of tau3-Bench at 8.04 against 7.63, 4.54 and 7.01.

Where it loses is the part listicles skip. Qwen3.6 scores 89.93 on PinchBench against 83.18, 37.62 on WebArena Verified against 26.75, and 83.69 on GPQA Diamond against 74.42. Gemma 4 also edges it on HLE at 8.57 against 7.40. So the honest summary is that LFM2.5 wins the tool-calling and multi-step agent tasks it was tuned for, and gives ground on browsing and knowledge-heavy reasoning.

Treat all of those numbers as vendor-reported, measured on Liquid AI's own harness, which is the model seller's preferred configuration. They are a starting point for your own evaluation, not a verdict, and the gap between benchmark scores and production behaviour is largest exactly where on-device agent claims are hardest to check.

What it costs to run locally: 14.53 GB at the small end

Memory is set by total parameters, not active ones, because every expert weight has to be resident even though only 2.45B are read per token. Liquid AI publishes six GGUF builds, and the file sizes from the repository's own file listing on 16 October 2026 are the clearest hardware guide available:

• QAD-Q4_0: 13.58 GB
• Q4_K_M: 14.53 GB
• Q5_K_M: 17.03 GB
• Q6_K: 19.69 GB
• Q8_0: 25.49 GB
• BF16 and F16: 47.96 GB

A 16 GB machine can technically load the Q4 file and then has almost nothing left for the operating system, the context cache or the agent process. Treat 32 GB of unified memory as the realistic floor for a comfortable local agent setup on Apple Silicon, and use the MLX 4-bit build on Macs, which is a separate download from the GGUF files. On a 24 GB or larger GPU, the Q5 and Q6 builds become the sensible default.

Speed claims are also vendor-reported: 184 tokens per second on an Apple M5 Max, up to 1.7x faster than the next-best comparable MoE on an AMD Ryzen AI Max+ 395 while using the least memory of the group, and up to 21,000 output tokens per second on a single NVIDIA B300 GPU when served at high concurrency. Native support covers llama.cpp, MLX, vLLM, SGLang, ONNX Runtime and LM Studio, so the practical question is which loader you already use.

The catch nobody quotes: the LFM Open License v1.0

The launch post says the models are open-weight and links to a model licence. The licence itself is where the commercial decision lives. The LFM Open License v1.0 defines a Threshold of 10 million US dollars in annual revenue, and Section 5 states that commercial-use rights are conditioned on you or your legal entity not exceeding it. Above that threshold, commercial use of the model or a derivative is simply not licensed under the agreement, with a carve-out only for qualified non-profit and research use.

That is not a defect in the model, and Liquid AI is entitled to set its terms. It is a procurement fact. If your company is past $10 million in revenue and you plan to ship a product with these weights inside it, you need a commercial agreement with Liquid AI before you build, not after.

The comparison matters because the alternatives are licensed differently. Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it are both tagged Apache-2.0 on their model cards (Gemma's card links to Google's own Gemma 4 licence page as well). NVIDIA's Nemotron-3.5-Lightning-30B-A3B-BF16 uses the OpenMDW-1.1 licence. If licence cleanliness outweighs the benchmark lead, Apache-2.0 is the safer pick, and it costs you roughly one benchmark category.

Local versus hosted: the per-token math

Liquid AI sells the same model through its own API. The pricing page lists LFM2.5-24B-A2B at $0.05 per million input tokens and $0.70 per million output tokens, with $2 of free credit for every new organisation and the same rate available through OpenRouter.

Put that next to a local deployment. At $0.70 per million output tokens, 10 million output tokens a month costs $7, and 100 million costs $70. A machine that comfortably runs the Q4 build, by contrast, is a capital purchase of several hundred dollars or more, plus electricity. Unless you are pushing serious volume, or your data cannot leave your network, the hosted endpoint is the cheaper way to find out whether the model is right for your workload.

Where local wins is on the two things the price list cannot show: marginal cost per token drops to zero once the hardware exists, and prompts, tool payloads and memory never leave the machine. For teams working with regulated data, or running agent loops that burn tens of millions of tokens overnight, that trade flips quickly. If you do not own the hardware and do not want to, renting GPU capacity by the hour is the middle path.

Who should run it, and who should not

LFM2.5-24B-A2B is a good buy for a narrow set of situations: you are building an agent that calls tools and runs multi-step tasks, you are under the licence revenue threshold, you have 32 GB of unified memory or a 24 GB GPU, and keeping prompts on-device is a requirement rather than a preference. The tool-use scores above are the reason to pick it over a denser model of similar size.

Skip it in these cases:

• You are past $10 million in annual revenue and shipping commercially. Take Apache-2.0 weights (Qwen3.6-35B-A3B or Gemma-4-26B-A4B-it) or negotiate with Liquid AI first.
• You need image input. LFM2.5 is text-only, while the Qwen and Gemma model cards list image-text-to-text as their pipeline task.
• You are on 8 GB or 16 GB of RAM. The smallest GGUF is 13.58 GB before runtime overhead; try a smaller LFM2.5 sibling instead of fighting the memory wall.
• Your workload is knowledge-heavy question answering. Liquid AI itself recommends pairing it with retrieval or web search rather than using it standalone.
• You need instant answers. The always-on thinking block adds tokens and latency to every request, including trivial ones.

At a glance

Open-weight local LLMs compared, all figures read from Hugging Face model cards and APIs on 16 October 2026

FAQ

Is LFM2.5-24B-A2B free for commercial use?

The weights are free to download, but the LFM Open License v1.0 conditions commercial-use rights on the licensee not exceeding $10 million in annual revenue (Section 5). Past that threshold the model is not licensed for commercial use, so you need a separate agreement with Liquid AI.

How much RAM does a 24B local LLM need?

Plan around the file size, since all 24B parameters have to sit in memory even though only 2.45B are active. Liquid AI's GGUF builds run from 13.58 GB for QAD-Q4_0 and 14.53 GB for Q4_K_M up to 25.49 GB for Q8_0, with the BF16 file at 47.96 GB. A 32 GB unified-memory machine is the realistic floor for comfortable local agent use.

LFM2.5-24B-A2B vs Qwen3.6-35B-A3B: which should you run?

On Liquid AI's own benchmarks LFM2.5 wins the tool-calling tasks (AutomationBench 27.69 vs 8.72, LiveClaw 48.09 vs 38.19) while Qwen3.6 wins PinchBench 89.93 vs 83.18, WebArena Verified 37.62 vs 26.75 and GPQA Diamond 83.69 vs 74.42. Qwen3.6 is also Apache-2.0 and accepts image input, so teams above the $10M revenue line usually start there.

Related reading

LM Studio vs Ollama for running local models, open-weight and sovereign AI models, Ollama alternatives for local inference, renting GPU capacity instead of buying hardware, What does Muse Spark pricing look like: API or your own Macs?

Sources

• Liquid AI launch post: LFM2.5-24B-A2B, 13 Oct 2026 (benchmarks, speeds, release)
• LFM2.5-24B-A2B model card (architecture, context length, training, guidance)
• LFM Open License v1.0 (Section 5 commercial-use threshold)
• LFM2.5-24B-A2B-GGUF repository (quantisation file sizes)
• Liquid AI pricing page (per-million-token rates for LFM2.5-24B-A2B)
• Qwen3.6-35B-A3B model card (Apache-2.0, 35.95B parameters, downloads)
• Gemma-4-26B-A4B-it model card (Apache-2.0 tag, 25.81B parameters)
• NVIDIA Nemotron-3.5-Lightning-30B-A3B-BF16 model card (OpenMDW-1.1 licence)
• Hacker News discussion, 14 Oct 2026 (285 points, 79 comments)