Skip to content

docs(blog): sync final local-llm-fast-enough article and fix stale benchmark claims#215

Merged
wesleyscholl merged 1 commit into
mainfrom
claude/squish-blog-sync-deploy-8ed169
Jul 7, 2026
Merged

docs(blog): sync final local-llm-fast-enough article and fix stale benchmark claims#215
wesleyscholl merged 1 commit into
mainfrom
claude/squish-blog-sync-deploy-8ed169

Conversation

@konjoinfinity

Copy link
Copy Markdown
Collaborator

Summary

  • Replace the stale, incomplete draft in docs/blog/posts/local-llm-fast-enough.md with the final, fact-checked article — this was live on squish.run before this PR, since docs.yml deploys on every push to main. Confirmed live: the "9.8x faster... on long prompts" framing and the disproven TTFT-loss claim were both fetched from https://squish.run/blog/local-llm-fast-enough/ before this change.
  • The new article adds the two architecture components the old draft was missing (the optimization engine, the memory governor) and corrects the headline speedup framing to 1.15–14.7x depending on prompt reuse.
  • Add the four mascot images the article references (docs/assets/blog/squish-hero.png, squish-building.png, squish-flying.png, squish-fistpump.png), placed following the site's existing assets/blog/ convention.
  • Fix the two stale claims already flagged in docs/paper.md:
    • The "~26s parity" cold-prefill claim → now states the real 1.15–1.32x Squish edge on cold/unique prompts.
    • The "192ms vs 167ms, Squish's only loss" TTFT claim → now states Squish leads TTFT at every tested length.
  • Fix a template bug in overrides/main.html: the front matter's date field is now a nested {created, updated} mapping, which was rendering as a raw Python dict inside the article:published_time OpenGraph meta tag. The template now emits clean article:published_time / article:modified_time tags for both the flat and nested date forms.

Type of change

  • Bug fix
  • New feature
  • Performance improvement
  • Documentation update
  • Refactor / cleanup

Checklist

  • pytest tests/ passes locally — N/A, no Python source changed
  • ruff check squish/ tests/ reports no errors — N/A, no Python source changed
  • No hardcoded absolute paths (no /Users/<name>/...)
  • No model weights, eval output files, or log files staged for commit
  • Changes are scoped to one logical concern (blog content sync + the two stale-claim fixes it surfaced)
  • Performance-sensitive changes include a before/after squish bench run — N/A, no perf-sensitive code changed

Verification

Built the site locally with the exact plugin versions docs.yml uses (mkdocs<2, mkdocs-material, mkdocs-exclude, mkdocs-rss-plugin) and confirmed via Playwright screenshots:

  • All 4 embedded SVG/CSS architecture diagrams render as live vector art, not escaped raw code.
  • All 4 mascot images resolve with correct relative paths and the floating hover animation is active.
  • No em-dashes, no old color values, corrected benchmark tables render correctly.
  • The Created/Updated date indicator renders correctly in the post sidebar.

Related issues

N/A — sprint task, no linked issue.


Generated by Claude Code

…nchmark claims

Replace the stale draft blog post with the fact-checked final article,
including the optimization-engine and memory-governor sections it was
missing and the corrected 1.15-14.7x framing (was: flat 9.8x claim).
Add the four mascot images the article references.

Fix two superseded claims in paper.md: the cold-prefill "parity" claim
is now the real 1.15-1.32x Squish edge, and the TTFT "only loss" claim
is now a Squish win at every tested length.

Fix overrides/main.html's OpenGraph template, which rendered a raw
Python dict into article:published_time when date front matter uses
the nested {created, updated} form.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants