What happened
Ollama’s model library now lists lfm2.5:24b, a package for Liquid AI’s LFM2.5-24B-A2B. The page was updated two weeks ago and showed 286K downloads for the lfm2.5 tag when the supplied source was captured. The library’s basic local invocation is ollama run lfm2.5:24b. It also provides examples for Ollama’s local chat API in cURL, Python, and JavaScript.
Liquid describes LFM2.5 as a family of hybrid models designed for on-device deployment. The 24B-A2B variant is described as the largest and final model in that family: a 24-billion-parameter reasoning model that activates about 2 billion parameters per token. Ollama lists the model architecture as lfm2moe. Its displayed Q4_K_M quantization is 15GB, while the package metadata names the LFM Open License v1.0. The listing gives the model a 128K-token context window.
The positioning is explicitly agentic. Liquid says the model was trained with reinforcement learning inside OpenClaw, Hermes Agent, Pi, and OpenCode. Ollama includes launch examples for OpenCode and Hermes Agent: ollama launch opencode --model lfm2.5:24b and ollama launch hermes --model lfm2.5:24b. The source says the model is intended for agents that reason and call tools, rather than presenting it as a general-purpose chat model alone.
The page also reports a Liquid-provided LiveClaw score of 48.09 for LFM2.5-24B-A2B, compared with 38.19 for Qwen3.6-35B-A3B. It characterizes the latter as having roughly 50% more parameters. This is a vendor claim on the library page, not an independent evaluation supplied with the source. Liquid further says the model “thinks before every answer” and frames its approximately 2B active parameters as making reasoning tokens cheaper.
The package supports English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish. Ollama advises pairing it with web search or retrieval for knowledge-heavy questions. The same library family also includes an 8B-A1B variant, but the material change here is the availability of the 24B-A2B package. The supplied verified source does not establish availability through Vercel AI Gateway, pricing there, or any other hosted endpoint.
What the archive adds
This is a new access event for a model family already covered in Liquid AI has published LFM2.5-24B-A2B-DSpark on Hugging Face. That earlier note concerned the DSpark draft model and reported throughput and concurrency claims; the new fact is a locally runnable Ollama package for the 24B-A2B reasoning model.
Trend signal
This modestly supports Harness architecture outlives model choice. The notable detail is not only the model’s weights but the supplied launch path into named agent harnesses. A model becomes easier to test operationally when the runtime can start it with the harness a team already uses.
Why now
The timing is only partly visible in the source: Ollama’s listing says it was updated two weeks ago. The archive shows that related Liquid model artifacts had already appeared, so this looks like a distribution step rather than evidence of a newly introduced model family.
Connections to other events
An earlier archive item described GGUF variants for LFM2.5-24B-A2B, including quantized files and llama.cpp commands. Ollama adds a different operational surface: a tag, local API examples, and launch commands for OpenCode and Hermes Agent. These routes can coexist, but they optimize different parts of adoption.
Why it matters
Teams evaluating local agents usually face two separate questions: whether a model can call tools reliably enough for their workflow, and how much work it takes to put that model behind an agent harness. This listing reduces the second question to a reproducible starting point. A team can run the documented package, attach a deliberately narrow tool set, and observe behavior in its own environment before committing to a hosted provider or a larger deployment.
The 15GB Q4_K_M package size is also a useful planning constraint rather than a performance guarantee. It signals that local use requires meaningful storage and appropriate hardware, while the sparse active-parameter claim does not by itself establish latency, throughput, quality, or total operating cost on a particular machine. Builders should treat the package as an evaluation option, not as proof that it will replace their current inference stack.
What happens next
- More agent harnesses will expose documented local launch paths for LFM2.5-24B-A2B. (3 months, 55%). Watch for: additional maintained commands or provider integrations.
- Teams testing the package will publish workload-specific tool-calling comparisons rather than relying on the listed LiveClaw score. (6 months, 60%). Watch for: reproducible harness configurations and task traces.
Limits and uncertainty
The supplied source is an Ollama library page and contains vendor-provided model descriptions and benchmark claims. It does not provide independent measurements of tool reliability, local latency, memory use beyond the listed 15GB quantization, or performance across hardware. It also does not state a release date for the underlying model. Although the story intake mentions Vercel AI Gateway, the verified primary material here is limited to Ollama; no hosted availability, pricing, rate limits, or provider behavior should be inferred from this note.
Skill to build now
Create a small local evaluation harness around ollama run lfm2.5:24b: choose three real tool-using tasks, restrict the model to the minimum required tools, save inputs and tool traces, and define pass conditions before running it. Add retrieval or search for tasks dependent on current or proprietary knowledge, as the listing recommends. Compare completed tasks, intervention rate, and wall-clock time rather than comparing model descriptions.
Reader takeaway
LFM2.5-24B-A2B now has a documented Ollama path for local, tool-using agent experiments, including OpenCode and Hermes Agent launch commands. That makes it straightforward to test as a private-agent candidate, but the available source does not yet prove its claimed benchmark advantage or establish a hosted Vercel offering.
Open question
How reliably does the local Ollama package complete real multi-step tool tasks under a team’s own permission, retrieval, and review constraints?