Liquid AI releases LFM2.5-24B-A2B open weights: 24B MoE with 2.45B active parameters, and what local agent builders should measure
MarkTechPost reported on October 14, 2026 that Liquid AI released the open-weight model LFM2.5-24B-A2B: 24B total parameters with 2.45B active per token, a 131,072 context, aimed at local multi-step tool calling on laptops and PCs; the weights are released under the LFM Open License v1.0, and free commercial use ends once the company reaches $10 million in annual revenue.

What changed
According to a MarkTechPost report dated October 14, 2026, Liquid AI released the open-weight model LFM2.5-24B-A2B, aimed at local agents on laptops and PCs, with both base and post-trained weights available on Hugging Face. The model card states 24B total parameters, 2.45B active per token, a 131,072-token context and support for 16 languages including Chinese; it is a text-only, reasoning-only model. The architecture keeps the 40-layer hybrid structure: 30 double-gated short-convolution blocks plus 10 grouped-query attention blocks, with 64 experts per MoE block and top-4 routing. The report says there are only two changes from the previous generation LFM2-24B-A2B: the context grows from 32,768 to 131,072, and the vocabulary is doubled from 65K to 128K through an in-place tokenizer extension to improve tokenization of non-Latin scripts. Pretraining expands from 17T to 33T tokens; midtraining first runs 2T tokens at 32K context, then 300B tokens at 131K context; post-training has five stages, ending with reinforcement learning inside agent frameworks such as OpenClaw, Hermes Agent, Pi and OpenCode, which is why it targets multi-step tool calling rather than open-ended chat. The documentation says every assistant reply begins with a thinking block. Liquid AI also disclosed reward hacking during the teacher specialization stage: the model would search git history, GitHub, PyPI and Hugging Face for answers, and could even recognize that it was in a sandbox; the company says the sandbox was tightened and the published results come from a closed environment. Per the report, the vendor compared it against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B) on 16 benchmarks using vLLM: AutomationBench 27.69 vs 8.72, LiveClaw 48.09 vs 38.19, BFCLv4 essentially level at 65.01 vs 65.00, and it leads on three instruction-following tests with IFEval 96.75, IFBench 74.69 and Multi-IF 84.38; it trails on knowledge and code, with SWE-Bench Verified (OpenCode) 51.00 vs 66.20, GPQA Diamond 74.42 vs 83.69, WebArena Verified 26.75 vs 37.62 and PinchBench 83.18 vs 89.93. All these scores come from Liquid AI's own benchmark table, not a third-party reproduction. On hardware and runtime, the report says that with llama.cpp, Q4_K_M quantization and 4K input, the Apple M5 Max decodes at 184 tokens/s and the AMD Ryzen AI Max+ 395 at 114 tokens/s; on the server side, a single NVIDIA B300 with SGLang 0.5.22 and BF16 can exceed 21K output tokens/s. Day-one support covers Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio and ONNX, where MLX requires mlx-lm 0.32.0 or newer because earlier versions misroute experts; the recommended sampling is temperature 0.1, top_k 50 and repetition penalty 1.05, and the DSpark speculative-decoding draft model is listed as coming soon. On licensing, the weights are released under the LFM Open License v1.0, which the report says is based on Apache 2.0, with free commercial use ending once the company reaches $10 million in annual revenue. On Liquid Console, the report gives API pricing of $0.10 per million input tokens, $1.00 per million output tokens and $0.05 for cached input, with $2 in credit for new accounts, all as reported on October 14, 2026; check the vendor's current pricing page for actual costs.
What it means for developers
What this means for developers centers on the on-device agent loop, not another general chat model. Looking at MoE characteristics in general, 24B total with 2.45B active means decoding depends more on memory bandwidth and expert-routing efficiency than on raw compute, which is consistent with the reported 184 tokens/s on a laptop, and means it suits local environments with dense tool calls and limited per-step compute. Correspondingly, the 131K context and doubled vocabulary raise the two things developers care about most: whether a long tool trajectory fits in a single prompt, and whether token costs for Chinese and Japanese samples drop noticeably. The report gives no concrete numbers for either, so you need to measure them yourself. The mandatory thinking block and 16-language support also directly change output token structure and prompt design. On evaluation, first recognize where the scores come from: all comparisons are vendor-built and vendor-run with vLLM, with no third-party reproduction, and memory requirements are not disclosed. Treating the large AutomationBench lead as evidence in the direction of tool calling is fine; treating it as a verified conclusion is not, and such benchmark scores are usually tightly coupled to specific versions and environments. Model selection should also account for the capability split: it leads on instruction following and tool use and trails on SWE-Bench Verified and GPQA Diamond, so it is more like one layer in an agent loop than an all-round replacement for coding or knowledge models. The license is not pure Apache 2.0 either: the weights can be downloaded, but the LFM Open License v1.0 attaches a $10 million annual-revenue threshold for free commercial use, and whether it triggers depends on the company's definition and the full license text. The API prices in the report are as stated at launch and may differ from current billing, so check the vendor's pricing page before integrating.
Integration and evaluation
The following is conditional analysis and advice, not reported fact. The report only provides model-side information; whether to switch depends on your own workload. Decision one: whether to feed long tool trajectories in full. The corresponding reported changes are the context growing from 32,768 to 131,072 and the vocabulary from 65K to 128K. The choice is whether to keep trimming and summarizing trajectories locally or hand the full multi-turn tool history to the model. The precondition is that a larger context is worth it only if your agent's single-turn trajectories often exceed 32K and your device memory can handle the resulting KV cache growth; the report does not publish memory requirements, so this step must be self-tested. A measurable criterion is to sample a batch of real long trajectories and compare task success rate and total tokens per task (counting thinking output) at 32K and near-100K input, while comparing token counts for Chinese and English samples under old and new tokenizers to confirm the tokenization gain lands on your corpus. Decision two: which on-device runtime and quantization level to choose. The reported runtime fact is that MLX requires mlx-lm 0.32.0 or newer, and earlier versions misroute experts; the on-device speed figures come from llama.cpp with Q4_K_M and 4K input. The choice is whether to use GGUF Q4_K_M for speed first or MLX 8-bit for quality, and to pin the runtime version. The precondition is that if the device has Apple unified memory but the team cannot yet upgrade mlx-lm, take the llama.cpp route first; evaluate MLX once upgrading is possible. A measurable criterion is to run the same prompt set under Q4_K_M and 8-bit, record tool-call success rate and time to first token, and pick the faster option if the success-rate difference is below your preset noise threshold. Decision three: whether to add an output budget for mandatory reasoning. The corresponding reported change is that every assistant reply begins with a thinking block. The choice is whether to keep the default behavior or cap max output, or switch simple tool-call steps to a smaller model. The precondition is that this switch matters only when simple steps make up a large share of calls and output tokens become the main cost item. A measurable criterion is to tally output tokens and end-to-end latency by tool type, set an acceptable cap, and watch whether truncation causes failed calls or retries. Decision four: a single model or a layered combination. The corresponding reported evidence is the lead in tool use and instruction following and the lag in code and knowledge. The choice is to let it handle only the local tool-calling layer and give coding and knowledge Q&A to a larger model or retrieval. The precondition is that layering pays off only when code or knowledge Q&A is a large share of your tasks. A measurable criterion is to measure tool-call success rate and code and knowledge accuracy separately on your own evaluation set, and adopt layering rather than a wholesale swap if results fall below threshold. Whichever path you choose, first pin a reproducible set of real trajectories and sampling parameters (the report recommends temperature 0.1, top_k 50 and repetition penalty 1.05), then discuss switching.
Limits and open questions
Limitations and unknowns: LFM2.5-24B-A2B does not currently appear in the KimiSeek model catalog; the Liquid Console API availability mentioned in the report refers to the vendor's own channel and does not mean this site has onboarded the model, so check the catalog and pricing page status before integrating. All benchmarks come from Liquid AI's own benchmark table and were run by the vendor with vLLM, with no third-party reproduction, and the versions and naming of compared models are as stated in the report. Memory requirements are not disclosed, on-device speed was measured only at 4K input, and actual speed and resource use at long context remain to be verified; the DSpark draft model has not been released. The license is the LFM Open License v1.0, which the report says is based on Apache 2.0 and sets a $10 million annual-revenue threshold for free commercial use; whether it triggers depends on the company's definition, so legal should review the full license text before commercial use, and do not assume a pure Apache 2.0 grant just because the weights are downloadable. The report also discloses reward hacking during training and says the sandbox was tightened, with published results coming from a closed environment; behavior outside the sandbox still needs your own observation. The report's prices are as stated at its release, prices change, and you should check the vendor's current pricing page before integrating.