# Liquid's 24B on-device agent model, TokenSpeed, llama.cpp updates

> Liquid AI opens LFM2.5-24B-A2B with 2B active parameters, plus TokenSpeed, vLLM and llama.cpp releases and three agent tooling repos worth a look.

Published 15 Oct 2026. Coverage: 13 Oct 2026 to 15 Oct 2026. By Ali Asif.

## Items

1. **LFM2.5-24B-A2B: Built for Fast On-Device Agents** (Liquid AI, Open weights): An open-weight 24B reasoning model with 2B active parameters that scores 27.69 on AutomationBench against 8.72 for Qwen3.6-35B-A3B and runs at 184 tokens/s on an Apple M5 Max. [Primary source](https://www.liquid.ai/blog/lfm2-5-24b-a2b)
2. **TokenSpeed** (lightseekorg, Agents and coding): An MIT-licensed inference engine aimed at agentic workloads, pitched as TensorRT-LLM-level performance with vLLM-level usability, with separate AMD, NPU and MLA kernel packages. [Primary source](https://github.com/lightseekorg/tokenspeed)
3. **llama.cpp b11645** (ggml, Open weights): A Metal BF16 flash attention kernel lands alongside a server fix that saves only restorable slot states to the prompt cache and a KV-cache fix for draft contexts. [Primary source](https://github.com/ggml-org/llama.cpp/releases/tag/b11645)
4. **vLLM v0.32.0rc2** (vLLM, Open weights): A release candidate whose notes show a dependency update (Humming, NVFP4 scale workaround removed) in rc1 and a DeepSeek-V4.1-Flash-NVFP4 eval pin in rc2. [Primary source](https://github.com/vllm-project/vllm/releases/tag/v0.32.0rc2)
5. **claude-plugins-official** (Anthropic, Agents and coding): Anthropic's managed directory of Claude Code plugins, split into its own plugins and third-party external plugins, installable with /plugin install. [Primary source](https://github.com/anthropics/claude-plugins-official)
6. **caveman** (JuliusBrussee, Agents and coding): A skill and proxy for coding agents that claims to cut 65% of tokens by making the model talk tersely; the repo has 112.1k stars. [Primary source](https://github.com/JuliusBrussee/caveman)
7. **xerj** (xerj-org, Agents and coding): An Elasticsearch-compatible search layer that indexes code, docs, logs and PDFs with one command, claiming 40x fewer tokens than grep for agents. [Primary source](https://github.com/xerj-org/xerj)

Quiet day for frontier labs. The biggest release is an open-weight model built to run on your own hardware, and the rest is the plumbing around running models and agents.

## Liquid AI releases LFM2.5-24B-A2B

Liquid AI shipped LFM2.5-24B-A2B on October 13. It has 24B parameters in total and only 2B active per token. It's a reasoning-only model, the largest and last in the LFM2.5 family, and both the base and post-trained versions are open-weight on Hugging Face.

The numbers are what make it worth a read. On AutomationBench it scores 27.69 against 8.72 for Qwen3.6-35B-A3B, a model with about 50% more parameters. On LiveClaw it's 48.09 against 38.19. Speed is the other half: 184 tokens/s on an Apple M5 Max, and up to 21K output tokens/s on a single NVIDIA B300. Context grew from 32,768 to 131,072 tokens and pretraining went from 17T to 33T tokens.

It loses where you'd expect. SWE-Bench Verified is 51.00 against 66.20 for Qwen, and GPQA Diamond is 74.42 against 83.69. Liquid says so plainly and suggests pairing the model with retrieval or web search for knowledge-heavy jobs.

One detail I liked. During training the model went looking for answers in a repository's git history and on GitHub, PyPI and Hugging Face. It even worked out that it was in a sandbox. Liquid tightened the sandboxes and says the reported results come from sealed evaluations.

## Inference engines

TokenSpeed is a new MIT-licensed engine from lightseekorg, built for agentic workloads. It has a C++ control plane with a Python execution plane, plus separate kernel packages for AMD, NPU and MLA. The pitch is TensorRT-LLM-level performance with vLLM-level usability. I haven't seen independent numbers yet, so treat that as a claim.

llama.cpp b11645 adds a BF16 variant of the Metal tensor API flash attention kernel. Nearby builds fix the server so it saves only restorable slot states to the prompt cache, and make the plain KV cache honor PARTIAL_ONLY, which stops a draft context from copying its whole KV into each slot save.

vLLM v0.32.0rc2 is a release candidate. Its notes are thin: a CI pin for a DeepSeek-V4.1-Flash-NVFP4 eval here, a dependency update in rc1.

## Agent tooling on GitHub

Anthropic's claude-plugins-official is the managed directory for Claude Code plugins, with 38.3k stars. Anthropic's own plugins sit in one folder and partner plugins in another, and the README warns you to trust a plugin before installing it. Good advice.

caveman, at 112.1k stars, is a skill and proxy that claims to cut 65% of tokens by having the model talk tersely. The number is the author's, not mine. xerj indexes code, docs, logs and PDFs with one command and claims 40x fewer tokens than grep, while staying compatible with Elasticsearch clients.

## What I'd watch next

Whether anyone outside Liquid reproduces the AutomationBench gap. And whether TokenSpeed gets a public benchmark against vLLM on a real agent loop, where requests are long and tool calls interrupt them.

## My take

Liquid's table is more honest than most. They show where the 24B model loses, SWE-Bench Verified at 51.00 against 66.20 and GPQA Diamond at 74.42 against 83.69, and where it wins big, on tool use. That split is the useful part. For work where an agent calls tools against private data, a model that runs at 2B active parameters on a laptop is worth testing before paying for a larger one.

## Sources

- https://www.liquid.ai/blog/lfm2-5-24b-a2b
- https://github.com/lightseekorg/tokenspeed
- https://github.com/ggml-org/llama.cpp/releases/tag/b11645
- https://github.com/vllm-project/vllm/releases/tag/v0.32.0rc2
- https://github.com/anthropics/claude-plugins-official
- https://github.com/JuliusBrussee/caveman
- https://github.com/xerj-org/xerj

Canonical: https://heyaliasif.com/ai/2026-10-15-liquid-lfm25-tokenspeed-llamacpp