v1.0.0 — autotune production release

What's new in v1.0.0

autotune proof — 30-second TTFT proof

Quick head-to-head that runs two scenarios: a short-prompt baseline and a long-context KV-allocation test. Forces a full KV buffer flush before both conditions to isolate KV init cost from model load time. Shows breakdown of load_ms (KV alloc) vs prefill_ms (prompt eval).

autotune proof --model qwen3:8b

autotune upgrade — in-place version management

Checks PyPI for the latest version and upgrades in place. A throttled upgrade reminder (once per 24 h) also prints after every CLI command.

autotune recommend — exact install commands in output

Each recommendation panel now shows the exact ollama pull command and autotune chat invocation. A 3-step Next Steps footer is appended to every run.

Model registry expanded to 40 models

Added smollm2-1.7b, phi3.5-3.8b, qwen2.5-3b, gemma3-4b/12b/27b, qwen2.5-7b, qwen2.5-coder-7b, deepseek-r1-8b, mistral-small-24b, qwen2.5-coder-32b, deepseek-r1-32b, llama-3.3-70b, deepseek-r1-70b. All models now include an ollama_tag for direct pull commands.

Bug fixes

• Model fitness honesty: TIGHT class introduced — SWAP_RISK no longer hard-blocks. Clear run vs chat distinction.
• autotune recommend pager: removed — output prints inline, works with pipe/redirect.
• Port conflict handling: autotune serve prints an actionable message instead of crashing.
• Telemetry race conditions: consent prompt now fires before SESSION_START on first run.

Install

pip install llm-autotune

Full changelog: https://github.com/tanavc1/local-llm-autotune/blob/main/CHANGELOG.md