Audit Together AI models and reasoning controls

[rekram1-node]

Summary

• audit every Together AI serverless chat model against the official provider catalog and deprecation schedule
• add 9 missing active models, yielding complete coverage of all 22 currently listed serverless chat models
• mark 7 past-removal endpoints deprecated and reconcile strongly established reasoning metadata
• add provider-specific reasoning controls only where Together documents them

Evidence

• Catalog and capabilities: https://docs.together.ai/docs/serverless/models
• Reasoning model types and controls: https://docs.together.ai/docs/inference/chat/reasoning
• Removal dates: https://docs.together.ai/docs/deprecations
• DeepSeek V4 Pro exact effort values: https://docs.together.ai/docs/deepseek-v4-quickstart

Reasoning semantics

• MiniMax M2.7: fixed reasoning (reasoning_options = [])
• Hybrid models: toggle only when Together explicitly documents reasoning={enabled: ...}
• GPT-OSS 20B/120B: exact low, medium, high effort values
• DeepSeek V4 Pro: toggle plus exact accepted high, max effort values
• Nemotron: toggle is represented; its documented medium/high control uses chat_template_kwargs.medium_effort, which does not map to the generic effort schema
• Qwen3 235B Instruct remains reasoning-capable without options because Together recommends its reasoning mode but does not document a formal request control
• models with reasoning = false omit reasoning_options

Coverage

• 29 provider files audited: 22 active, 7 deprecated
• 13 active reasoning models, 9 active non-reasoning models
• active controls: 1 fixed, 9 toggle-capable, 3 effort-capable; DeepSeek V4 Pro contributes to both toggle and effort counts

Verification

• bun validate
• provider consistency count/check script
• git diff --check
• no Together API credential was available, so no live inference tests were run