Pocket Specialist: fine-tune LFM3-1.7B into a 1GB offline agent for your CLI's own tools
Turns a tool's docs into synthetic training tasks and a 1GB local agent model. Paid by makers of developer tools.
Competitive · LFM3-1.7B agentic, 1GB at 4-bit
Written for: Solo builder
Next step
This week, generate 2,000 synthetic tasks for one popular CLI (gh or ffmpeg), LoRA-tune LFM3-1.7B, and publish the before/after pass rates. Pitch the result to that tool's maintainers on GitHub Discussions.
Why now
A tool-calling model small enough to ship inside a CLI or a browser (Opera's WebAI standard points the same way) used to be too weak for anything beyond toy tasks. An agent-trained 1.7B base plus a published environment-synthesis method makes narrow, offline assistants practical. Few dev-tool makers have the ML skills to do the fine-tune themselves.
What became possible
Liquid AI's LFM3-1.7B is post-trained for agent harnesses (Hermes Agent, Pi, OpenClaw), has a 256K context window, needs about 1GB at 4-bit, and is designed for quick domain adaptation. The SynthAPI paper shows how controlling the diversity of synthetic tools, tasks and databases shapes tool-use training, which is a recipe for generating the fine-tuning data.
Two ways in
- pocket-tune
CLI that reads an OpenAPI spec or --help output, generates SynthAPI-style tasks and mock environments, LoRA-tunes LFM3-1.7B, and exports a GGUF plus a pass-rate report against the untuned base.
- Customer
- Maintainers of open-source CLIs and small dev-tool companies that want an offline 'ask' command
- Model
- Open-source pipeline. Hosted GPU training runs at $49 each
- Docs-in-a-model
The same pipeline pointed at a product's docs and API, producing a model that runs in the visitor's browser.
- Customer
- Small SaaS with an API and support load
- Model
- $99 per build plus rebuilds when docs change
Competition
The browser Docs-in-a-model variant looks emptier. I found WebGPU runtimes but no one fine-tuning a SaaS's docs and API into a model that runs in the visitor's browser. It would compete for the same budget as server-side docs chatbots, though, and it has to justify a roughly 1GB download and weaker answers on low-end devices. Offline use, privacy, or zero inference cost would have to matter to the buyer, which is a narrower pitch.
Full competition notes (10 competitors)
The core pipeline is already taken: synthetic tool tasks, a small-model LoRA, GGUF export, and an eval against the base model. Liquid AI's own LQH harness and cookbook, distil labs, DeepFabric, Needle and NeMo all cover it, and Unsloth makes training on LFM models free. The biggest risk is that Liquid itself ships an OpenAPI or --help importer in LQH and wipes out the LFM-specific angle. What I didn't find is a turnkey, maker-facing product built around one specific CLI. That would mean parsing --help and OpenAPI with no hand-written seed examples, generating sandboxed mock environments so pass rates come from actual execution rather than an LLM judge, and shipping a drop-in 'mytool ask' subcommand plus a model that can be redistributed. The model would also need retraining in CI on every release, so it doesn't drift from the tool's flags. If you sell to maintainers, that release-time retrain and the published pass-rate badge are probably the defensible part, not the training itself. Open-source maintainers rarely pay, so the realistic buyers are small dev-tool companies. The browser Docs-in-a-model variant looks emptier. I found WebGPU runtimes but no one fine-tuning a SaaS's docs and API into a model that runs in the visitor's browser. It would compete for the same budget as server-side docs chatbots, though, and it has to justify a roughly 1GB download and weaker answers on low-end devices. Offline use, privacy, or zero inference cost would have to matter to the buyer, which is a narrower pitch.
- LQH (Liquid Harness) by Liquid AI
This is the closest overlap, and it comes from the company that makes the base model. It is a terminal agent that takes a task description, writes a synthetic data pipeline, scores samples with an LLM judge, runs baselines, and fine-tunes a Liquid Foundation Model (LFM2-1.2B-Instruct by default) into a task-specific on-device model. It is in closed beta and is general-purpose. It does not start from an OpenAPI spec or --help output.
- Liquid AI cookbook: home-assistant tool-calling example
Liquid's official recipe for a local tool-calling agent on llama.cpp. The steps are: build a benchmark taxonomy, generate synthetic data, fine-tune, and deploy back into the app. It is free, documented know-how that does a large part of what pocket-tune would do.
- distil labs
A startup platform that fine-tunes small models from a prompt and a few dozen examples, including for function calling. It handles synthetic data generation, training and evaluation, and you can download the model to run locally. It claims a 0.6B specialist beat its 120B teacher on an IoT function-calling task (vendor's own number). It gave PyCon DE 2026 talks on small-model tool calling.
- Making bad CLIs fun with Small Language Models (PyCon DE 2026 talk)
A talk describing synthetic data generation for any CLI, used to fine-tune Gemma 3 270M into a natural-language-to-command model that runs without API keys or a GPU. That is almost exactly the pocket-tune idea, presented as a technique rather than a product.
- DeepFabric
Open source (always-further/deepfabric). It takes an MCP server or a set of tools, generates tool-calling datasets from real executed traces, and fine-tunes small models. The author says a tuned Qwen3-4B beat frontier models on the Blender MCP server.
- Cactus Compute Needle
Its generate-data CLI turns a tool schema JSON file into synthetic tool-calling training examples for small on-device models.
- NVIDIA NeMo tool-calling tutorial
An official workflow: synthetic tool-calling data via Data Designer, LoRA fine-tuning of Llama 3.2 1B, then evaluation. It is tied to NVIDIA's platform rather than being a drop-in CLI.
- Unsloth LFM2.5 fine-tuning notebooks
Free Colab SFT and GRPO LoRA notebooks for LFM2.5-1.2B with GGUF export. This makes the training and export step a commodity.
- Gorilla CLI
A Berkeley tool that turns plain English into shell commands using hosted LLMs. It shows demand for an 'ask' layer over CLIs, but it is not offline and not specific to one tool.
- WebLLM / in-browser LLM chatbots
The runtime layer for the Docs-in-a-model idea. It runs quantized 1 to 3B models in the browser on WebGPU with offline caching. I found no product that fine-tunes a SaaS's own docs and API into a browser model.
Checked 2026-10-28
Risks
- Check Liquid's model licence terms before selling commercial fine-tunes
- A 1.7B model may still fail multi-step tasks often enough to annoy users
- Buyer demand is unproven. Most tool makers may prefer a hosted API call
Strong team fit; market pull is the weak point (4).
tech 6 · pull 4 · space 7 · fit 8 · build 6
Momentum of its main topic
Sources
Status history
- 2026-10-28 Flagged as buildable