[martinezharo]

Summary

Switch AI title generation from @cf/meta/llama-3.2-3b-instruct to @cf/google/gemma-4-26b-a4b-it.

Benchmark (16 snippets, same prompt, same sampling)

Snippet	Llama 3.2 3B	Gemma 4 26B-A4B
TSX UserCard component	UserCardComponent	UserCard
Java EventEmitter class	emit-event	EventEmitter
TS retry w/ backoff	retryFunction	retryWithBackoff
ESLint JSON config	next-config	eslint-config
GitHub Actions YAML	ci-deploy	github-actions-ci
Go health handler	handleHealthResponse	health-check-handler
Spanish relative-time fn	format_date	relativeTime
SQL orders per user	user_orders	user-order-count

The rest were equivalent (useDebounce, fetch_user_data, define-port, red-background, deploy-checklist, square_even_numbers).

Cost / latency

• Llama 3.2 3B: $0.051/M in, $0.335/M out → ~$0.000024 per title
• Gemma 4 26B-A4B: $0.10/M in, $0.30/M out → ~$0.000045 per title (~430 in / ~5 out tokens)
• Latency typically 400–700 ms vs ~250 ms (occasional spikes seen). Fine for a background title.

Notes

• Gemma reasons by default and spent all 24 max_tokens thinking (empty output) → chat_template_kwargs.enable_thinking: false.
• Gemma returns an OpenAI-style choices[0].message.content payload (confirmed through the real env.AI binding), so the route reads that and falls back to response.

Test plan

• tsc, eslint, vitest pass
• Verified the binding response shape with a temporary worker using remote AI

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	klipcode	40853d2	Commit Preview URL Branch Preview URL	Oct 20 2026, 10:37 AM

[oli-review]

Review

Switches snippet-title generation from the Llama 3.2 3B model to Gemma 4 26B-A4B, disables Gemma's default thinking budget, and adds a normalizer that accepts either the legacy { response } or OpenAI-style { choices[0].message.content } Workers AI payload before passing the text to the existing sanitizer. No other routes or callers change.

Risk: low

---

Automated review — opencode-go/deepseek-v4.1-flash · 0 comment(s)