Model · Liquid AI via Hugging Face
LFM2.5: 2B active parameters for an agent that thinks first
On October 13, 2026, Liquid AI introduced LFM2.5-24B-A2B: a local model with 24B total parameters, 2B active parameters, mandatory reasoning and tool calls. Its interest lies in the architecture: selected experts limit computation per token, while memory requirements and time to a completed result need separate assessment.
What this means for a buyer
This model merits consideration for assistants that obtain information through tools and act on the results. Buyers need to connect its MoE architecture to an actual workflow. A small active portion of the weights does not, by itself, determine server size or total task duration.
What a small active portion means
According to Liquid AI’s announcement, the model retains its predecessor LFM2-24B-A2B’s architecture: 40 layers and 64 experts per MoE block, with four selected experts. MoE means mixture of experts—model components sharing the processing work. The selected portion handles each token. Thus, 2B active parameters describe the computational path, while 24B total parameters describe the complete parameter set.
This is particularly interesting for agents because of their intermediate steps. An agent may receive command output, interpret it and choose another call before composing an answer. Computation repeats for every generated token, including reasoning. The smaller active path provides an architectural reason to investigate this model for extended tool workflows. It does not establish how many simultaneous agents a particular machine can serve.
2B active parameters do not make the weights 2B
The system still needs somewhere to store the required weights and supply selected experts for computation. Weight placement and numerical precision determine requirements separately from the active path. The published Transformers example loads the model in bfloat16. That is a specific execution option, not a description that automatically applies to another precision or every possible local configuration.
Long conversations add another memory demand: cached attention state, commonly called the KV cache. It supports processing continuations and depends on the operating configuration. Server selection should therefore account separately for weights, this cache and concurrent requests. The A2B label helps explain computation, but it cannot replace a description of how the chosen runtime places the model and serves multiple tasks.
Reasoning comes before the answer
LFM2.5-24B-A2B is presented as reasoning-only: it produces an explicit reasoning process before its final answer. This directly affects the interface. Token-generation speed and the wait for a useful result become different quantities. Users may wait for intermediate thinking to finish, while a tool agent also waits for commands to execute. Short operations therefore need an explicit decision about acceptable waiting time.
Context grows from 32K to 131K tokens, and tokenizer vocabulary from 65K to 128K; the authors connect the vocabulary expansion to non-Latin text. Vocabulary determines how text is split into elements, rather than adding another context window. The window itself does not establish useful recall of every fact: history, tool results and generated continuation share the budget. An entire repository is not automatically the best input for investigating one defect.
Connect the agent through a local runtime
The announcement lists support for llama.cpp, MLX, vLLM, SGLang and ONNX. Its proposed arrangement serves the model locally through an OpenAI-compatible backend, which the agent harness contacts. Compatibility here concerns the message format; computation can remain local. The Hermes example separately configures the server address, model and tool use.
This helps separate model quality from harness behavior. Reproducing a defect requires available tools and correctly returned call results. Announced runtime support does not establish that your complete combination works. Liquid AI also publishes LFM2.5-24B-A2B-Base. The ready agent described in the announcement calls for the corresponding agent-trained model; the base variant does not automatically inherit its properties. The weights use LFM Open License v1.0. Section 5 restricts commercial use by reference to a US$10 million annual-revenue threshold; companies need to account for those terms when selecting the model.
Choose by the complete sequence of actions
Liquid AI separately discusses decoding on Apple M5 Max and aggregate token output at high concurrency on one NVIDIA B300. These are different workloads: one concerns a local decoding stream, the other overall server output. The announcement does not disclose complete precision, input and batching configurations for these measurements. Their numerical speeds therefore cannot become a performance promise for a purchased machine.
A short trial can reproduce a known defect: obtain a log through a tool, select the relevant file, run a check and explain the result. Observe completion of the entire sequence and elapsed time, including reasoning. LFM2.5’s specific opportunity is serving as a local executor of tool workflows in an existing harness. Its architecture warrants that trial; the meaningful result is a completed action sequence that demonstrates suitability for the team.
What to check before choosing
- What happens when a command returns partial output, an access error or an unexpected format?
- Can tool-call history be retained separately from reasoning text to investigate failed tasks?
- Which operations require a user decision before writing to a repository or external system?
- Does tokenization preserve Russian paths, identifiers and mixed-language names in working materials?
Buyer implications are our editorial interpretation. The publication does not establish the speed or compatibility of a particular catalogue system.