fix(ollama-cloud): correct 3 context_lengths, add 3 served-but-missing tags

[mikkihugo]

All values read from the provider via ollama.com/api/show, the same source the existing deepseek-v4-flash:0731 entry already cites.

1. Corrected limit.context (3 models)

model	current	provider reports
glm-5.2	976,000	1,000,000
minimax-m3	512,000	524,288
nemotron-3-nano:30b	1,048,576	262,144

nemotron-3-nano:30b is the one that bites: the recorded window is 4x the real one, so a client sizing requests from this database overruns the model and the provider rejects the request. I hit exactly that in practice before checking against the API.

2. Added 3 tags that are served but not catalogued (new files)

GET https://ollama.com/v1/models lists these three; the database had no entry.

model	context
deepseek-v4-pro:0813	1,048,576
deepseek-v4-pro:preview	524,288 — half the stable tag
deepseek-v4-flash:preview	1,048,576

How to reproduce

curl -s https://ollama.com/api/show \
  -H "Authorization: Bearer $OLLAMA_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"model\":\"nemotron-3-nano:30b\"}" \
  | jq -r ".model_info | to_entries[] | select(.key|endswith(\"context_length\"))"

Observed:

glm5.2.context_length            = 1000000
minimax-m3.context_length        =  524288
nemotron-3-nano.context_length   =  262144
deepseek4.context_length         = 1048576   # deepseek-v4-pro:0813
deepseek4.context_length         =  524288   # deepseek-v4-pro:preview
deepseek4.context_length         = 1048576   # deepseek-v4-flash:preview

I checked all 17 other ollama-cloud entries the same way; they already match and are untouched.

Scope notes

• Only providers/ollama-cloud/ is changed.
• minimax-m3 is deliberately not aligned with the MiniMax entries. minimax/MiniMax-M3 and minimax-coding-plan/MiniMax-M3 are both 1,000,000, and Ollama genuinely serves a smaller window for the same model — the per-provider split here is correct, only the ollama-cloud number was off.
• output on the corrected models is untouched: api/show reports no max output, so I have no provider-sourced value. On the three new files output mirrors context, following every other deepseek entry under this provider; that field is convention, not API-verified.

[github-actions]

Action items

• [high] [violation] providers/ollama-cloud/models/glm-5.2.toml:12 - Check: Override-only base_model files must not restate identical inherited values. Why: After this change, limit.context = 1_000_000 matches models/zhipuai/glm-5.2.toml exactly, so the provider [limit] block is a redundant override (merge blocker). Action: Remove the entire [limit] section so context/output inherit from zhipuai/glm-5.2.
• [high] [violation] providers/ollama-cloud/models/nemotron-3-nano:30b.toml:13 - Check: Override-only base_model files must not restate identical inherited values. Why: limit.context = 262144 now matches models/nvidia/nemotron-3-nano-30b-a3b.toml (262_144); only output = 131072 still differs from the lab’s 262_144. Action: Drop context from [limit] and keep only the real output delta (output = 131072).

[Copilot]

Pull request overview

Updates the ollama-cloud provider catalog to match the actual served context windows reported by Ollama’s api/show, preventing clients from constructing requests that exceed the provider-enforced limits.

Changes:

• Correct limit.context for glm-5.2 to 1_000_000.
• Correct limit.context for minimax-m3 to 524_288.
• Correct limit.context for nemotron-3-nano:30b to 262_144.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
providers/ollama-cloud/models/nemotron-3-nano:30b.toml	Reduce limit.context to the provider-reported 262,144 tokens to prevent request overruns.
providers/ollama-cloud/models/minimax-m3.toml	Adjust limit.context to 524,288 to match api/show.
providers/ollama-cloud/models/glm-5.2.toml	Increase limit.context to 1,000,000 to match api/show.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Action items

• [high] [violation] providers/ollama-cloud/models/deepseek-v4-pro:0813.toml:1 - Check: Non-lab hosts must use base_model; provider files stay override-only. Why: Ollama Cloud did not create DeepSeek V4; lab metadata already exists at models/deepseek/deepseek-v4-pro-0813.toml, and the peer deepseek-v4-flash:0731.toml correctly uses base_model. This full inline definition restates lab fields and skips inheritance (including lab open_weights = false, which this file sets to true). Action: Rewrite as override-only with base_model = "deepseek/deepseek-v4-pro-0813", keep only host deltas (name, reasoning_options, [limit] overrides, optional status), and drop duplicated identical fields.
• [high] [violation] providers/ollama-cloud/models/deepseek-v4-flash:preview.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable. Why: This is a host tag of DeepSeek V4 Flash, not an Ollama-origin model. Full inline definition omits required resolved fields (release_date, last_updated) that lab metadata would supply. Action: Point base_model at deepseek/deepseek-v4-flash (or the dated lab snapshot if that is what :preview serves), keep only Ollama-specific overrides (name, reasoning_options, context/output limits, optional status = "beta"), and do not restate unchanged lab capabilities/modalities.
• [high] [violation] providers/ollama-cloud/models/deepseek-v4-pro:preview.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable. Why: Same third-party host rule as above for DeepSeek V4 Pro; full inline definition again omits release_date/last_updated and duplicates lab-shaped fields. Action: Use base_model = "deepseek/deepseek-v4-pro" (unless a distinct lab snapshot is justified), keep only real host deltas including the half-size [limit] and optional status = "beta".
• [medium] [violation] providers/ollama-cloud/models/deepseek-v4-flash:preview.toml:7 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: All three new files set reasoning_options with toggle but only document catalog/API context sources, not the Ollama request field. Peer deepseek-v4-flash:0731.toml documents think / reasoning_effort. Action: On each new DeepSeek file, add a leading comment matching the established ollama-cloud peer (e.g. native think = true|false; OpenAI path reasoning_effort / reasoning.effort = high|max).