Together AI launches canary rollouts: metric gates catch 137% p95 regression at 10% traffic, auto-rollback
zainhas · x · 2026-09-23
Together AI shipped Canary Rollouts for its dedicated inference offering, letting teams upgrade production models without downtime.
- Live traffic moves from the current model to a new checkpoint in gated steps, with health checks before each step
- Canary stages support metric gates (e.g. p95 latency, error rate) evaluated after each step; a tripped gate pauses the rollout at the current traffic share
- Teams can cancel and reverse the rollout while live requests keep flowing
A real case: a Qwen2.5-7B → Qwen3.5-9B canary caught a 137% p95 regression at just 10% of traffic and was canceled and reversed with zero failed requests. Docs, CLI reference, and a rollout API were released alongside.
Related event: Together AI Launches Canary Rollouts for Zero-Downtime Model Upgrades(2 posts)→
More from coding & agent
- Combos CLI launches: AI agents generate game assets and publish H5 games — lxfater · 2026-09-23
- Dev builds online 3D ski game with GPT-6 Sol, all assets and backend via Combos CLI — lxfater · 2026-09-23
- The first Muse UI element ever designed was the agent status pill — alexandr_wang · 2026-09-23
- Claude Opus 5.5 writes 35.5x kernel on KernelBench-Mega, beating GPT-6 Astra's 24.8x — scaling01 · 2026-09-23
- KVMem pages KV state to give agents million-token workspaces on a consumer GPU — rohanpaul_ai · 2026-09-23
- Researcher finds 26 vulnerabilities in 19 AI coding agents, including 12 RCEs and MCP flaws — matthew_d_green · 2026-09-23