Liquid AI's laptop-size agent model is a routing decision, not a replacement

*The tool-calling half of an agent can now run on hardware you own; the knowledge half cannot, and a new audit says check what those agents install.*

Opening read

Liquid AI's new LFM2.5-24B-A2B is the largest step yet in a run of the company's claims that agent tool calling can run on hardware you own, and this time the vendor's own numbers are specific enough to test. The same post concedes that larger models still win the knowledge-heavy tests, so the decision for IT leaders is which steps to route home, not which vendor to drop. Further down, Trail of Bits has counted how badly package managers check what they download, which matters more the moment agents run on machines you control.

On-device agents: Liquid's 24B model moves the tool-calling slice home

The open question for anyone running AI models on their own computers has been whether a small local model can be trusted with multi-step tool use, the kind of work where an agent calls a piece of software, reads the result, and calls again. The settled view is sensible: small open models (ones you can download and run yourself) are fine for drafts and sorting, but anything that must call tools reliably across many steps belongs on a frontier model (a lab's most capable tier) rented through a vendor's API (the paid online connection an application uses to call it). Liquid AI has been pushing against that view since August. Back then it pitched its 2.6-billion-parameter model as competitive with models four times its size on tool use and multi-step agent tasks, and its 3-billion vision model as level with two rival small models, Gemma-4-E2B and Qwen3.5-2B, on tool-use tests Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm2-5-2-6b) Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm2-5-vl-3b). Tuesday's 24B release is the biggest version of the same claim, and the one with the most usable numbers. The model scores 48.09 on LiveClaw, a test of agent tool use, against 38.19 for Qwen3.6-35B-A3B, a rival open model, and by Liquid's own count it comes out ahead or even on five of the eight tool-use tests it ran, among them BFCLv4 and ToolSandbox. The catch sits in the same post. Liquid concedes that larger models still hold the edge on knowledge-heavy tests such as GPQA Diamond, HLE, and AA-Omniscience, which measure what a model knows rather than what it can do Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b).

The speed figures turn this into a hardware question. Liquid reports 184 tokens (the word fragments a model reads and produces) per second on an Apple M5 Max device and up to 21,000 per second on a single Nvidia B300 server chip, and in August its 2.6B model managed 220 per second on the same class of M5 Max chip while the 3B vision model hit 228 Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b) Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm2-5-2-6b) Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm2-5-vl-3b). A model roughly nine times larger giving up about a sixth of the speed is the number that matters most here, and I will not guess at the design reason behind it. What it means for the routing choice is simpler: if the vendor figure holds, moving tool steps onto one high-end machine need not mean accepting a much slower agent. The hardware in question is a top-end Apple chip, not whatever sits on every desk in the fleet.

My departure from the consensus is not that frontier models are finished for agents. It is that an agent is not one workload. Some of its steps are plumbing: pick a tool, pass the arguments, read the result, repeat. Others need the model to recall something it was never handed. Liquid's results say the plumbing can run locally on the tests it chose and the recall still trails, and if most of your agent's steps turn out to be plumbing, that becomes a routing decision for the engineer building the system and a spend decision for the leader paying for API calls on steps that never needed a frontier model. That split is a hypothesis to test on your own traces (the logged step-by-step record of what your agent did), not a settled fact about agent work. Keep the knowledge steps in the cloud and the sensitive data on the box until the traces say otherwise.

I hold that read loosely. Every figure above is Liquid's own, run on Liquid's chosen comparisons, and I apply my standing rule that a vendor benchmark predicts nothing about your traces until you run them. The local stack is also still moving under you. Ollama's latest release candidate (a test build ahead of a stable version) ports Apple-specific kernels, the low-level code that runs the math on a chip, to CUDA, Nvidia's chip software, and llama.cpp, another free local-model runner, just fixed a crash where machines reporting zero free memory produced invalid layer splits and aborted loading Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.5-rc0) llama.cpp Releases (https://github.com/ggml-org/llama.cpp/releases/tag/b11626). That maintenance is part of the price.

So test it. Pull one day of traces from an agent you already run and sort each step into two bins: steps that call a tool or format a result, and steps that require the model to recall a fact unaided. Run both bins through the Liquid model on the hardware you would actually use and through your current cloud model, and count completed steps in each. A matching count in the first bin does not settle anything on its own; it marks that slice as a candidate for a closer look at cost, reliability, and who maintains the box. A gap that shows up only in the second bin is consistent with the weakness Liquid already admits, not a reason to keep paying cloud prices for plumbing. A gap in both bins means the vendor numbers did not travel, and you have lost one afternoon.

Supply chain: most package managers will not catch a tampered download

Local agents install code, and Trail of Bits has shown that the tools doing the installing mostly take it on faith. The firm pointed 25 package managers (the tools that pull code libraries into a project) at a compromised repository, and most of them would not catch a swapped file by default, because fewer than half ship with a lockfile (the record of the exact versions a project expects) that checks for tampering, and only two of the 25 checked every single download they pulled Trail of Bits (https://blog.trailofbits.com/2026/10/14/package-managers-have-an-integrity-problem/). August's ChainDrop worm showed the cost. A self-spreading worm that began with one hijacked maintainer account reached more than 400 packages on npm, the main public store of JavaScript code libraries Microsoft Security Blog (https://www.microsoft.com/en-us/security/blog/2026/08/04/chaindrop-supply-chain-compromise-anatomy-self-propagating-worm/) Wiz Research (https://www.wiz.io/blog/keyv-and-cacheable-npm-supply-chain-attack). The worm showed what a single tampered release can do, and the audit shows most installers would not notice it. For any engineer letting a coding agent run installs, the rule is simple: pin (lock every dependency to an exact version), verify, or do not let it fetch.

Earth AI: planet-scale models become agents you can call

Climate AI is getting easier to reach on two fronts at once: access and simulation cost. On access, Google is widening experimental use of its Geospatial Reasoning agent through Ask Google Earth and an Agent-to-Agent service, a protocol that lets your own agent hand a question to Google's, and the agent lets Gemini query Google's weather, flood, population, and satellite models alongside its mapping and public datasets Google Research Blog (https://research.google/blog/putting-earth-ai-to-work-geospatial-agents-for-our-planets-biggest-challenges/). On cost, Ai2's HiRO-ACEv2 simulates rare rain, wind, and pressure extremes at 3 km resolution for about 1/300th the cost of the X-SHiELD atmospheric model Allen Institute for AI (https://allenai.org/blog/hiro-acev2). Neither is a price list. Google's access is still experimental, which makes it a test option rather than a vendor a business leader can put in a budget, and Ai2's figure compares the cost of running a simulation, not what climate-risk analytics will sell for. For the engineer, A2A access is the path to test now; for the leader, the signal to wait for is general availability.

Election traffic: Anthropic puts a number on political chat

Last week Anthropic rewrote the rules for what customers may do with Claude in influence operations and surveillance, and this week it published how much political use there actually is. From 250,000 conversations over one week, Anthropic found that roughly one in seventeen US conversations, about 5.75 percent, touches politics, public policy, or social issues, while just over half a percent of all Claude conversations, 0.51 percent, deal with elections or voting as such. Claude already handles that narrow slice with a specific behavior, a banner that sends voters to TurboVote, and that is the useful part for anyone outside Anthropic: the company has a stated, election-specific response ready for the 0.51 percent Anthropic News (https://www.anthropic.com/news/2026-usage-policy-update) Anthropic News (https://www.anthropic.com/news/2026-us-midterms) Princeton CITP - Freedom to Tinker (https://blog.citp.princeton.edu/2026/10/08/ai-election-policy/). Any business leader shipping a Claude-based assistant to US users needs the same kind of stated policy for those turns before November, not after.

Augur calls

• Call: Google will announce general availability of its Geospatial Reasoning agent, replacing experimental access, by 2027-03-31.
• Criteria: A Google Research, Google Cloud, or Google Earth blog post or product page stating the Geospatial Reasoning agent is generally available to developers or enterprises; experimental, preview, or waitlist status counts as a miss.
• Due: 2027-03-31
• Odds: 30%

What to watch

• Liquid AI's five-of-eight lead-or-tie on tool-use tests is self-reported; independent reruns of LiveClaw and BFCLv4 on the 24B-A2B model will show whether the on-device number travels. Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b)
• Ollama's CUDA port of its Apple-chip kernels is still a release candidate; a stable 0.40.5 would bring the ported Apple kernels to Nvidia machines. Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.5-rc0)
• Trail of Bits found only two package managers that verify every download; the next move is which ecosystems turn integrity checks on by default. Trail of Bits (https://blog.trailofbits.com/2026/10/14/package-managers-have-an-integrity-problem/)
• Google calls Geospatial Reasoning access experimental; the Augur call above tracks whether that becomes general availability by March. Google Research Blog (https://research.google/blog/putting-earth-ai-to-work-geospatial-agents-for-our-planets-biggest-challenges/)
• Anthropic's 0.51 percent election share is a mid-October baseline; a follow-up count closer to the vote would show whether queries climb. Anthropic News (https://www.anthropic.com/news/2026-us-midterms)

Verification notes

Every benchmark and speed figure for LFM2.5-24B-A2B comes from Liquid AI's own post and is unaudited in the evidence, and the LiveClaw comparison is against a single named rival Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b). The August figures for Liquid's smaller models are likewise the vendor's, and the evidence names the M5 Max as a chip or device, not a specific machine, so the speed comparison is across a chip class rather than one shared box. Google's geospatial agent remains experimental access rather than a shipped product, and Anthropic's percentages reflect one week of its own sample.

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-15?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-15)