fix: Mark 6 removed Together.ai serverless models as status = "deprecated"

[github-actions]

I made the data fix. Below is the review-ready PR description.

Summary

Issue #3324 asked to mark 6 Together.ai serverless models as status = "deprecated" because they are no longer served via the serverless endpoint. Verified each model against Together.ai's authoritative first-party serverless models catalog. Five of the six models are confirmed absent from the current serverless catalog and were marked deprecated. One model from the issue, Qwen/Qwen3.6-Plus, is still actively listed in Together.ai's serverless catalog (with pricing that exactly matches the existing TOML), so it was intentionally left active — the issue's claim for that model is contradicted by the authoritative source.

Changes

Added status = "deprecated" (placed after open_weights, following the serializer's field order) and a leading deprecation source comment to five files:

File	Change
providers/togetherai/models/zai-org/GLM-5.1.toml	Added status = "deprecated" + deprecation source comment
providers/togetherai/models/zai-org/GLM-5.toml	Added status = "deprecated" + deprecation source comment
providers/togetherai/models/Qwen/Qwen3-235B-A22B-Instruct-2507-tput.toml	Added status = "deprecated" + new leading source comment (file had no header)
providers/togetherai/models/Qwen/Qwen3.5-397B-A17B.toml	Added status = "deprecated" + deprecation source comment
providers/togetherai/models/essentialai/Rnj-1-Instruct.toml	Added status = "deprecated" + new leading source comment (file had no header)

No other fields (cost, limits, modalities, descriptions, release/last_updated dates) were modified. Qwen/Qwen3.6-Plus.toml was not touched.

Evidence

• Together.ai "Available models" / serverless catalog (https://docs.together.ai/docs/serverless-models, accessed 2026-07-18): The live first-party chat-models table lists the current serverless models. The five deprecated models above are absent from this table, confirming they are no longer offered via serverless. Conversely, Qwen/Qwen3.6-Plus is present in the table (context 1,000,000; input $0.50; output $3.00 — matching the existing TOML exactly), which is why it was left active.
• Schema definition (packages/core/src/schema.ts:246): status: z.enum(["alpha", "beta", "deprecated"]).optional() confirms status = "deprecated" is a valid top-level field.
• Existing convention: Already-deprecated Together.ai models (e.g. deepseek-ai/DeepSeek-V3.toml, moonshotai/Kimi-K2.5.toml, Qwen/Qwen3-Coder-Next-FP8.toml) use the same status = "deprecated" pattern, confirming the convention.

Validation

• Verified each of the six issue models against the live first-party serverless catalog page (fetched 2026-07-18) by name.
• Confirmed status is a schema-valid optional top-level field via packages/core/src/schema.ts.
• Confirmed togetherai is not a synced provider (not in packages/core/src/sync/index.ts provider map), so these hand-authored files are not rewritten by the daily sync; leading comment blocks are preserved.
• Re-read all five edited files to confirm correct TOML structure and comment placement at the top of each file.
• I did not run bun validate (no shell/Bash access in this environment); the workflow's validation step will confirm parse/schema correctness.

Review notes

• Qwen/Qwen3.6-Plus discrepancy: The issue lists it as removed, but Together.ai's serverless docs (accessed 2026-07-18) still list Qwen/Qwen3.6-Plus as an active serverless chat model with pricing identical to our TOML. I did not deprecate it. If the reporter's curl errors were due to a transient issue, key/access scoping, or a docs page lagging behind an imminent removal, a follow-up can deprecate it once the first-party catalog no longer lists it.
• Descriptions intentionally preserved: Existing deprecated Together.ai models use description = "Legacy model retained for compatibility with older integrations". I deliberately did not rewrite the five models' descriptions, because (a) the issue only requests the status change, (b) that boilerplate is an editorial convention rather than a verifiable factual claim (and may be inaccurate — these models may be fully gone, not "retained"), and (c) replacing them would discard the original descriptive text. Flagging in case a maintainer prefers the uniform description convention applied.
• Out-of-scope models not addressed: While cross-referencing, I noticed two other Together.ai models are also absent from the current serverless catalog and are not already deprecated: meta-llama/Meta-Llama-3-8B-Instruct-Lite and LiquidAI/LFM2-24B-A2B (the catalog now lists LiquidAI/LFM2.5-8B-A1B instead). These were intentionally not modified because they are outside this issue's scope; mentioning them here as a potential follow-up.
• essentialai/Rnj-1-Instruct is absent from the serverless catalog and was deprecated as requested; I did not investigate whether the model name itself is canonical, as that is outside this issue's scope.

Closes #3324

Automated by the issue fixer: https://github.com/anomalyco/models.dev/actions/runs/29632028261

[github-actions]

Action items

• [high] [violation] providers/togetherai/models/zai-org/GLM-5.toml (and the four other deleted files: zai-org/GLM-5.1.toml, Qwen/Qwen3.5-397B-A17B.toml, Qwen/Qwen3-235B-A22B-Instruct-2507-tput.toml, essentialai/Rnj-1-Instruct.toml) - Check: Previously-served provider models that are no longer listed must be marked status = "deprecated", not deleted. Why: The README schema reference defines deprecated as "the model is no longer served by the provider's public API" — exactly this scenario — and the repo uses it 100+ times, including 7 existing togetherai entries (e.g. Qwen/Qwen3-Coder-Next-FP8.toml:8, deepseek-ai/DeepSeek-V3.toml:8). Deleting discards pricing, limit, modality, and capability history that the catalog deliberately preserves, and breaks external consumers that reference these model IDs. The PR's own title and body state the intent is to "mark ... as status = \"deprecated\"" and that "No other fields ... were modified," directly contradicting the second commit's full deletion of all five files. Action: Restore the five files with status = "deprecated" and the leading deprecation comment as the first commit did, instead of deleting them. If deletion is genuinely intended, rewrite the PR title and body to justify overriding both the issue's deprecation request and the documented convention.

[github-actions]

No actionable findings.