Building a Browser-Native Research Agent with LFM2.5 and WebGPU

🤷 Nico Martin 1w Liquid AI just released LFM2.5-2.6B, and it is wild what a model this compact can do. I built a research agent that runs entirely in the browser and puts its full think, plan, and act loop on display. Give it a question and it: - Creates an explicit action plan - Reasons about the next step - Chooses and calls tools - Inspects each result - Updates the plan and repeats It does not stop after one tool call. The agent keeps looping until every step in the plan is complete. Along the way, it can ask clarifying questions, use local context, delegate focused research to an isolated Wikipedia subagent, and collect sources. Once the plan is done, it turns the evidence into a cited response and downloadable research paper. The incredible part is the size: just 2.6B parameters. That makes LFM2.5 seriously fast on-device with WebGPU, while still being capable enough to plan, reason, use tools, and finish multi-step objectives. The model, agent loop, and inference pipeline all run locally. No inference server, no account, and prompts and outputs stay in the browser. Small models are not just autocomplete. They can be agents at the edge. Try the research agent in your browser: https://lnkd.in/dZwa9PAN Explore the code on GitHub: https://lnkd.in/d4aq-zEE More about the release: https://lnkd.in/dqXb5dvk

Confirmed Accurate

• "Liquid AI just released LFM2.5-2.6B" — Confirmed. Liquid AI released LFM2.5-2.6B (and a base checkpoint) on August 4, 2026, an on-device agentic model with open weights on Hugging Face.
• "Just 2.6B parameters" — Accurate. Official specs put it at 2.69B total parameters (30 layers: 22 double-gated short-convolution blocks + 8 GQA blocks).
• "Capable enough to plan, reason, use tools, and finish multi-step objectives" — Consistent with Liquid AI's published benchmarks, which show LFM2.5-2.6B leading instruction-following and most tool-use benchmarks against larger models (Gemma-4-E2B/E4B, Qwen3.5-4B/9B), trailing Qwen3.5-9B only on BFCLv4.
• General on-device/WebGPU framing — Accurate at the architecture level. Liquid AI itself ships an official WebGPU browser demo for this exact model ("LFM2.5 Edge Research Agent" on Hugging Face Spaces), so browser-based, local, tool-using agents built on this model are technically plausible and already demonstrated by the vendor.

Mostly Accurate

• "Seriously fast on-device with WebGPU" — Directionally true, but the specific throughput numbers being circulated (220 tok/s, 30 tok/s on phone) come from Liquid AI's native benchmark environment (Apple M5 Max, mobile CPU), not necessarily from WebGPU-in-browser execution specifically. Browser/WebGPU throughput is typically lower than native inference due to overhead; the post implies WebGPU speed without citing browser-specific numbers.
• "No inference server, no account, prompts and outputs stay in the browser" — Plausible and consistent with how WebGPU-based Transformers.js/ONNX-Web demos work generally, but this is a claim about Nico Martin's specific implementation that cannot be independently verified without inspecting the linked code.

Partially Accurate

• "Small models are not just autocomplete. They can be agents at the edge." — This is a fair generalization but overstates uniformity. Liquid AI's own documentation explicitly does not recommend LFM2.5-2.6B for "agentic coding or knowledge-heavy tasks," and notes it trails larger models meaningfully on coding benchmarks (LiveCodeBench: 59.41 vs. 69.86 for Qwen3.5-9B). A "research agent" arguably falls into knowledge-heavy territory, which is a tension the post doesn't address.

Not Fully Verified

• The specific agent architecture described ("explicit action plan," "isolated Wikipedia subagent," "downloadable research paper," "clarifying questions") — These are first-person claims about a personal project. They are plausible given the model's advertised tool-calling and planning capabilities, but none can be verified without testing the linked demo or reviewing the linked GitHub code directly.
• Real-world reliability of the agent loop — Liquid AI's own release blog notes that in at least one agentic test run, the model pulled data via an MCP tool, misformatted it, and confidently presented incorrect output as fact — a hallucination risk in multi-step agentic pipelines. Whether Nico Martin's implementation has guardrails against this is unstated.

Opinion / Promotional Language

• "It is wild what a model this compact can do" — subjective enthusiasm.
• "The incredible part is the size" — promotional framing.
• "Small models are not just autocomplete. They can be agents at the edge." — engagement-bait-style closing statement, presented as a broader truth rather than a claim specific to this model/benchmark set.

Missing Context

• Benchmarks are Liquid AI's own. The "beats 9B models" framing traces back to Liquid AI's self-reported comparisons (ToolSandbox, Multi-IF, IFStruct), not independent third-party evaluation. Self-reported benchmarks should be treated with normal skepticism about test selection and methodology.
• Vendor's own use-case caveat. Liquid AI explicitly steers this model away from "knowledge-heavy tasks," which sits in tension with building a "research agent" on it — not mentioned in the post.
• No mention of hallucination/reliability risk in agentic tool-use chains, which Liquid AI's own release notes flagged as an observed failure mode.
• No performance disclosure for the browser demo itself (load time, model quantization, memory footprint in-browser, device requirements for WebGPU support) — relevant since browser-based LLM inference is highly hardware- and browser-version-dependent (Liquid's own audio WebGPU demo notes minimum browser versions, e.g., Edge 113+).