feat(together-ai): update model YAMLs [bot]

[models-bot]

Auto-generated by poc-agent for provider together-ai.

---

Note

Medium Risk
Incorrect pricing or retirement flags could misroute clients or underestimate cost; widespread removeParams and modality edits affect request validation for multimodal APIs.

Overview
Bulk sync of Together AI provider model metadata to match current pricing, capabilities, lifecycle, and API parameter behavior.

Pricing and limits are updated across several chat and multimodal models (e.g. Qwen3.7-Max, DeepSeek V4 Pro variants, Seedream 5.0 lite, google/flash-image-3.1-lite with per-image output cost, google/veo-3.1 with per-request cost). Context windows and output limits change for models such as LFM2.5-8B, Qwen3.8, GLM-5.2, and gpt-oss-120b.

Lifecycle marks multiple models deprecated or retired with dates (Qwen3-30B-A3B-Base, DeepSeek-R1-Original, FLUX.1-schnell, cogito-v2-1-671b, multilingual-e5-large-instruct, gpt-oss-20b). Qwen3.5-4B-Lora is set active; nemotron-3-nano-omni moves to preview.

Capabilities and params: video/image models gain audio (or image) modalities where applicable; gemma-4-E2B-it drops video input. Many non-chat models add removeParams: reasoning_effort. Prism-ML/Ternary-Bonsai-27B gains configurable reasoning_effort; DeepSeek-V4-Pro-0813 adds low as a supported value. Minor sources URL updates accompany several entries.

^{Reviewed by Cursor Bugbot for commit d44f142. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

/test-models

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d44f142. Configure here.}

[cursor]

Off-by-one GLM-5.2-FP8 context window

Low Severity

context_window is set to 1048575 (2^20 - 1) instead of the documented 1048576 token window. Sibling GLM-5.2-FP8-Lora and other providers' GLM-5.2 entries use 1048576, so this catalog value can disagree with the live limit by one token.

 

^{Reviewed by Cursor Bugbot for commit d44f142. Configure here.}

[cursor]

GLM-5.2-FP8 drops reasoning_effort

Medium Severity

This thinking chat model now lists reasoning_effort in removeParams, so callers cannot set thinking effort. Sibling GLM-5.2 exposes reasoning_effort with high/max (default max), and GLM-5.2 accepts that parameter. Stripping it hides a supported control on the FP8 serverless variant.

 

^{Reviewed by Cursor Bugbot for commit d44f142. Configure here.}