optimize-local-model-compression

A guide for compressing local AI models, especially models used for calling tools, by reducing their numerical precision. It focuses on choosing compression methods and settings that preserve structured output such as valid JSON and reliable tool calls.

In plain words
What is it for?
Use it when comparing 4-bit quantization methods, converting a model for local use, investigating broken JSON or lost tool calls, or comparing a compressed model with its hosted version.
Why use it?
Ordinary model compression can save memory while making tool calls fail or produce malformed JSON. This guide helps diagnose that trade-off and choose a suitable method.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/understudylabs/understudy-agent-tools/optimize-local-model-compression
Any agent
npx skills add understudylabs/understudy-agent-tools --skill optimize-local-model-compression
Clone the repo
git clone --depth 1 https://github.com/understudylabs/understudy-agent-tools

Made for: Claude Code, Codex.

Per session 98 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 2,256 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00098 $0.02256
Opus 5 $0.00049 $0.01128
Sonnet 5 $0.00020 $0.00451
Haiku 4.5 $0.00010 $0.00226

Measured 2d ago against content hash 5e879513c115, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

optimize-local-model-compression scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/optimize-local-model-compression/SKILL.md · 186 lines

How it starts

The opening of the file, as written. The whole thing — 186 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Optimize Local Model Compression

Compression (quantization to 4-bit) is mandatory for local deployment — it is the difference between a 9.5 GB model and a 3.6 GB model, between 23 tok/s and 45 tok/s. But standard compression methods optimize for general text quality (perplexity, MMLU) and silently destroy tool-call fidelity. This skill teaches the method we invented to fix that: outcome-optimized, layer-aware compression that protects the circuits responsible for structured output.

This is not theory. Every recommendation below is backed by measured results on Gemma 4 E2B through 31B, on a 103-row tool-call benchmark and a 14-task multi-turn agent board, on Apple Silicon (M5 Max, 128 GB), at zero API cost.

When to use

  • The user's compressed model emits broken JSON, fails to call tools, or scores lower on tool-calling tasks than the hosted version.
  • The user is choosing between quantization methods (naive 4-bit, QAT, OptiQ, or a custom conversion).
  • The user wants to convert a model and is asking which settings matter.
  • A local eval shows a model "can't do tool calls" but the hosted version can.

Safety Gates

  • No conversion without explicit approval. Conversions are compute-heavy (15-30 min GPU time) and produce large artifacts. State the target model, BPW, and expected output size first.
  • Background long conversions. Sensitivity probing takes 15-30 min; background it and keep working.
  • Do not run conversions during meetings. GPU-bound work can freeze external monitors on Apple Silicon. Check for active display connections before starting.

The three things that matter

1. Group size must match the training noise profile

If the model is a QAT checkpoint (Google's q4_0 format), the downstream quantization group_size must match the QAT block structure (32). The default in most frameworks (MLX: 64, llama.cpp: varies) silently breaks QAT's training-time hardening.

Setting Group size Tool-name accuracy Errors
QAT g64 (broken) 64 (default) 0.369 19
QAT g32 (correct) 32 (matched) 0.379 8

Read the full file on GitHub · 186 lines

Files

What ships with it

2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 186 lines · 98 tokens per session scan A 5e879513c115

Subscribe to this mod's changes

optimize-local-model-compression is a skill published in the GitHub repository understudylabs/understudy-agent-tools (16 stars, last pushed 3d ago), licensed MIT. It adds 98 tokens to every session and 2,256 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

systematic-debugging

Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes.

obra/superpowers · 21 tokens

brainstorming

You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.

obra/superpowers · 37 tokens

auto-perf-optimize

Run agent-driven VS Code performance or memory investigations. Use when asked to launch Code OSS, automate a VS Code scenario, run the Chat memory smoke runner, capture renderer heap snapshots, take workflow screenshots, compare run summaries, or drive a repeatable scenario before heap-snapshot analysis.

microsoft/vscode · 62 tokens

chat-perf

Run chat perf benchmarks and memory leak checks against the local dev build or any published VS Code version. Use when investigating chat rendering regressions, validating perf-sensitive changes to chat UI, or checking for memory leaks in the chat response pipeline.

microsoft/vscode · 51 tokens

chat-pet-sprite-creation

Use when creating or changing VS Code chat pet sprite art, sprite sheets, state animations, eye treatments, Stable/Insiders variants, or pet transitions under src/vs/workbench/contrib/chat/browser/widget/media/chatPet.

microsoft/vscode · 53 tokens

cpu-profile-analysis

Analyze V8/Chrome CPU profiles (.cpuprofile) and DevTools trace files (Trace-.json). Use when: profiling performance, investigating slow functions, comparing code paths, finding bottlenecks, analyzing timeToRequest, understanding call trees from sampling profiler data, analyzing layout/paint/rendering, investigating…

microsoft/vscode · 71 tokens