Together cuts Llama 3.3 70B and DeepSeek-V4 from serverless today
A migration wave lands today: Together drops two heavily-used serverless IDs while Gemini flags its Veo 3.1 and Omni previews. Grep for dead model names.
Copy markdown
Llama 3.3 70B Turbo leaves Together serverless
`meta-llama/Llama-3.3-70B-Instruct-Turbo` is removed from Together's serverless inference today (2026-10-22). Apps that hardcode it will error — switch to `meta-models/Muse-Glimmer-30B`, or keep 3.3 70B running on a dedicated endpoint, which Together still supports.
DeepSeek-V4-Flash is pulled too — bump to 4.1
`deepseek-ai/DeepSeek-V4-Flash-0731` also exits serverless the same day; Together points you to `deepseek-ai/DeepSeek-V4.1-Flash`. It's close to a drop-in, but re-run your evals — 4.1 Flash is a newer checkpoint, not just a rename.
Gemini flags its Veo 3.1 video previews for shutdown
All three Veo 3.1 preview IDs (`veo-3.1-generate-preview`, `-fast-`, and `-lite-`) are listed for an Oct 22 shutdown. If you generate video through the Gemini API, Google's single named successor is `gemini-omni-1.1-flash`.
Omni Flash preview folds into the same GA model
`gemini-omni-flash-preview` is deprecated alongside the Veo trio — all four route to `gemini-omni-1.1-flash`. Google calls these 'earliest possible' dates with advance notice, so you won't be cut off silently, but start the port now.
Why it bites: dead model IDs 404, they don't degrade
A removed model ID returns an error, not a quiet fallback. Together's two removals are firm for today; grep your code and configs for all six strings now, before a prod call to a retired model starts throwing.