Changelog

Local Quench 0.1.1

in preparation: what the next release brings so far

Fixes and the first piece of the shared leaderboard.

What's new

• Shared QuenchBench results. When you choose to share a benchmark run, it goes to the public leaderboard at localquench.com (numbers only: model, quantization, engine settings, graphics card, results per task). Suspicious runs are held back. You can delete everything you shared at any time in Settings > Data sharing.
• What quantization costs. A shared run also says how the model was served: the engine (LM Studio runtime, llama.cpp, Ollama, vLLM), the KV cache type and the sampling settings. Results with a different KV cache are kept apart, and the QuenchBench page shows, for each model run in several quantizations, how much each one solves against the largest one.
• Benchmarks from the command line. quenchbench.py in this repository downloads the app and runs QuenchBench, SWE-bench, Terminal-Bench or Aider Polyglot on any model your server serves. The Windows build now prints into the console it is started from.
• Start screen. Opening the app no longer drops you into the last chat or creates an empty one: it shows the start screen and asks which folder to work in. New chat asks the same.
• Tidier sidebar. Days, projects and benchmark runs start collapsed; the group of the chat you open expands by itself.
• Benchmark errors you can see. A run that fails or stops shows the reason right away, with a clear box in Settings > Benchmark.

Fixes

• One name everywhere: on Windows start LocalQuench.exe (local-forge.exe is still there for existing shortcuts), the Android app is LocalQuench.apk, and the command line calls the benchmark quenchbench. Settings, chats and results stay where they were.

• Models that call tools with other argument names (command instead of cmd, pattern instead of glob) no longer get stuck in a loop: the call just works, and a call that still doesn't fit says which arguments the tool takes.

• A git commit without a path no longer sweeps up untracked junk files.

• Weak models that lose track after a long chat keep their task list and their progress (which commands worked, which files they already read).

• The chat export shows when each message came and how long each reply took, and counts tokens on cloud models.

• Updates also come straight from this repository when localquench.com is unreachable.

Small models

• Tool calls written as text are understood in every common shape: Hermes and Qwen <tool_call> JSON (also a list, Python quoting, or cut off before the closing tag), GLM's <arg_key> form, Llama's <|python_tag|>, Mistral's [TOOL_CALLS], DeepSeek's markers, a fenced or bare {"name": …, "arguments": …} object and Gemma's tool_code blocks. Arguments with a raw newline, a trailing comma or Python quotes are repaired instead of refused.
• A tool set per model. The new default auto gives small and mid-size models the light set (15 tools: the essentials plus file outline, git and the web) and strong models every tool; mcp and load_skill appear only when something is configured for them. The edit tools are described for the edit format in use, and a 12B-or-smaller model writes whole files unless it was trained on an edit tool.
• Sampling presets from the model cards for Qwen3 and Qwen3.5, Qwen3-Coder and Coder-Next, Qwen2.5-Coder, Gemma 3/3n/4, gpt-oss, GLM-4.x, Devstral/Codestral/Mistral Small, Nemotron 3, DeepSeek-R1 and V3.x, Kimi K2, MiniMax M2, Llama 3/4, Phi-4 reasoning, SmolLM3 and Seed-OSS.
• Fewer wrong nudges. "Continue" only after a promised action ("Let me run the tests"), not after "I should note that…"; the own-checks reminder counts check runs, not reads, and treats a new module or test file as work on the project; the polling warning fires only for sleep loops; a repeated check_process while a build runs is not a loop; one harness note per tool result; after a take-stock or troubleshoot note the turn goes on instead of ending on the recap.
• Small windows. The reasoning cap, the self-review diff, file reads, command output and the compaction request all scale with the model's context; the reply's token limit never exceeds what is left of the window; a benchmark prompt never names a tool that is not offered.

QuenchBench

• Comparable rows only. The leaderboard groups runs by task selection as well as model, quantization, KV cache and card; runs with harness switches changed, the bare baseline, hand-picked task lists and runs under 13 tasks are kept out of it; a run on a cloud API is marked and shows no graphics card; memory is shown as a range.
• Benchmarks use the standard harness, never a profile the tuner adopted for the model, and the tuner adopts one only on a significant held-out result.
• Read-by-hand tasks (what does the program print) no longer allow running the program; answer tasks are graded on the prose of the final message, so pasting the file is not an answer; the sandbox hides the task files from commands; --container on the command line.

Local Quench 0.1.0 – first public beta

2026-10-22

A coding agent built around small local models.

Most coding agents are made for frontier models and leave a small model to fend for itself. Local Quench goes the other way: the harness around the model is designed for what small models actually do on real tasks, and helps them through it. It watches for the usual failures and steers the model back before they cost you an afternoon:

• it announces a step and stops without doing it,
• it keeps rewriting its own test script instead of the code,
• it sends an edit that no longer matches the file,
• it trusts make | tail when the build actually failed,
• it forgets what it already did once the chat gets long,
• it loops on the same call with the wrong arguments.

Every one of these rules came from transcripts of local models on real work, and each can be switched off and measured with QuenchBench, the benchmark built into the app. Frontier models (Claude, GPT, Gemini, through your own key) get a lighter harness that stays out of their way.

What's in it

• Agent with real tools: file edits with undo, shell commands in a sandbox, tests, git, a repo map, an in-app browser preview, sub-agents for big repositories.
• Made for local servers: LM Studio, llama.cpp, Ollama and vLLM, with context size, KV cache and thinking settings per model.
• QuenchBench: what your model on your card can be trusted with, by kind of work and language. Plus SWE-bench, Terminal-Bench and Aider Polyglot runs (beta).
• Model advisor: which models and quantizations fit your VRAM and how fast they will run.
• Phone and tablet: check on a long task or start a new one from your phone.
• Costs: what each chat would have cost on a cloud model, next to your electricity.

Download

Or from https://localquench.com/download.html. The app updates itself.

Good to know

• This is a beta. If something breaks, open an issue or use Settings > General > Report a problem in the app.
• Windows may show a SmartScreen warning and Android an "unknown developer" one: the builds are not signed with a paid certificate yet.
• Your code stays on your computer. Nothing is shared unless you turn it on, see the privacy policy.