Add Tempr's Cohere, AI21, Inception, Morph and Sarvam models

[chrissalomon]

Follow-up to #7638 (Tempr as a provider): the chat models five more labs' own APIs currently serve, 17 in total. Tempr addresses them as <lab>/<id> on the caller's own key for that lab.

Checked live on 2026-09-30 and again on 2026-10-13 (UTC), rebuilt on current dev:

• Cohere (10): each model id got a validation error for an out-of-range temperature on v2/chat (400 "temperature must be between 0 and 2.0"), which Cohere returns only after resolving the model. An unknown id gets 404 "model not found".
• Sarvam (1): GET /v1/models lists sarvam-105b, and it answers chat with reasoning_effort low, medium and high. xhigh is rejected ("Input should be 'low', 'medium' or 'high'").
• AI21, Inception, Morph (6): checked against each lab's docs the same day, since there's no key for them here. AI21's versioning table points jamba-large at Jamba Large 1.7 and jamba-mini at Jamba2 Mini (2026-01), with neither on its deprecation list. Inception's model list has mercury-2.5 and mercury-2, and no deprecation. Morph's API docs list morph-v3-fast and morph-v3-large; the only deprecation there is a router endpoint.
• All 17 reasoning_options match Tempr's GET /v1/models for the same ids.

Only models each lab currently serves (see above):

• Cohere (10): the Command A family (command-a-plus-05-2026, command-a-03-2025, -reasoning-08-2025, -vision-07-2025, -translate-08-2025), Command R and R+ (08-2024), Command R7B and R7B Arabic, and north-mini-code-1-0. None has a date on Cohere's deprecations page. Left out the Aya models, which have no price in providers/cohere.
• AI21 (2): jamba-large and jamba-mini. The older versioned ids are retired.
• Inception (2): mercury-2.5 and mercury-2. Left out mercury-edit-2, which Inception serves only on its FIM and edit endpoints, not chat completions.
• Morph (2): morph-v3-fast and morph-v3-large. Left out auto, Morph's router between the two, since it isn't a model of its own.
• Sarvam (1): sarvam-105b. Left out sarvam-30b: Sarvam's API answers "Model 'sarvam-30b' has been deprecated". No [cost], same as providers/sarvam: Sarvam prices in rupees. Sarvam switches thinking off with a JSON null effort, which Tempr doesn't send, so it lists effort low/medium/high (the values Sarvam's API accepts) with no off switch.

The entries: base_model points at each lab entry with the lab's own pricing, since it's the same model on the caller's own key. north-mini-code-1-0 is $0, as in providers/cohere. reasoning_options are what Tempr's GET /v1/models reports, and each file's leading comment lists the fields a caller sets. Reasoning docs: https://temprhq.io/docs/gateway-chat-completions#reasoning

The Inception and Morph entries use the models/inception and models/morph lab files that #9531 added; this PR no longer adds its own. bun validate passes with this branch merged onto current dev (2026-10-13).

🤖 Generated with Claude Code

[github-actions]

No actionable findings.

[chrissalomon]

@rekram1-node, a small bump on this one. I rebuilt it on today's dev: #9531 added the models/inception and models/morph lab files this PR was adding, so it now carries only the 17 Tempr entries and points at those. bun validate passes, and all 17 still answer on the labs' APIs today (description updated).

#7675 (OpenAI) also still merges cleanly, with the recheck in my comment there. If either needs changes, or you'd rather I split or batch them differently, just say and I'll do it.

[github-actions]

No actionable findings.