feat(ollama-cloud): add deepseek-v4-flash:0731 model

[renaudcerrato]

Context

Ollama Cloud has been serving deepseek-v4-flash:0731 since 2026-07-31 (the 0731 snapshot of DeepSeek V4 Flash, with enhanced agentic capabilities and integrated DSpark speculative decoding). The lab metadata already exists on dev at deepseek-v4-flash-0731.toml; this PR adds the missing provider entry so the model surfaces under ollama-cloud in the generated catalog.

Note: An earlier open PR, #3976, adds the same model. This PR is a corrected alternative — see "Relationship to #3976" below. Happy to defer to the maintainers on which to land.

Confirmed live on the Ollama Cloud API:

$ curl -s https://ollama.com/api/tags | jq -r '.models[] | .name' | grep deepseek-v4-flash
deepseek-v4-flash
deepseek-v4-flash:0731

$ curl -s https://ollama.com/api/show -d '{"model":"deepseek-v4-flash:0731"}' | jq '.'
{
  "capabilities": ["completion", "tools", "thinking"],
  "details": { "family": "deepseek4", "parameter_size": "304180418494", "quantization_level": "FP8" },
  "model_info": { "deepseek4.context_length": 1048576, "general.architecture": "deepseek4" },
  "modified_at": "2026-07-31T08:00:00-07:00"
}

What changed

• Added providers/ollama-cloud/models/deepseek-v4-flash:0731.toml, a base_model-factored provider entry that inherits provider-agnostic facts from the existing lab metadata and only declares host-specific overrides.

Critical Changes

• New model on ollama-cloud: deepseek-v4-flash:0731 now resolves in the generated catalog with limit.context = 1_048_576 / limit.output = 1_048_576 and reasoning_options of toggle + effort: ["high", "max"].
• Override-only: the provider file uses base_model = "deepseek/deepseek-v4-flash-0731" and restates only reasoning_options and [limit] (real deltas vs. the lab's 1M/384K limits). name, description, family, dates, capabilities, modalities, open_weights, license, weights, and benchmarks are inherited from the lab entry.
• Reasoning controls: toggle + effort: ["high", "max"] matches the DeepSeek V4 family on this provider (see deepseek-v4-flash.toml and deepseek-v4-pro.toml) and the AGENTS.md DeepSeek V4 guidance (toggle + high/max, not L/M/H).
• Context override rationale: the lab's 1_000_000 is a rounded figure; Ollama Cloud's /api/show exposes the real deepseek4.context_length = 1_048_576. output = context follows the established ollama-cloud convention (feat(ollama-cloud): add deepseek-v4-flash model #1586).

Expected behavior

• bun validate passes (exit 0).
• Resolved provider model:
  • id: deepseek-v4-flash:0731
  • name: DeepSeek V4 Flash 0731 (inherited from lab)
  • limit: { context: 1048576, output: 1048576 }
  • reasoning_options: [{ type: "toggle" }, { type: "effort", values: ["high", "max"] }]
  • All other fields inherited from models/deepseek/deepseek-v4-flash-0731.toml.

Relationship to #3976

#3976 covers the same model. This PR fixes three issues in that one:

1. Model ID matches the live API. feat(ollama-cloud): add deepseek-v4-flash:0731-cloud #3976 names the file deepseek-v4-flash:0731-cloud.toml (and overrides name to the same string). The Ollama Cloud API exposes deepseek-v4-flash:0731 — both /api/tags and /v1/models — with no -cloud/:cloud suffix. The suffixed form has been a recurring source of 400/404 errors in downstream consumers (e.g. NousResearch/hermes-agent#16179).
2. name inherited, not restated. Per AGENTS.md override-only, name should not be restated when it doesn't truly differ; if it does, it shouldn't be the raw ID. This PR inherits the lab's display name "DeepSeek V4 Flash 0731".
3. reasoning_options match the DeepSeek V4 siblings. AGENTS.md specifies toggle + high/max for DeepSeek V4 (no none). The existing deepseek-v4-flash.toml and deepseek-v4-pro.toml both use [{ toggle }, { effort: ["high","max"] }]. feat(ollama-cloud): add deepseek-v4-flash:0731-cloud #3976 drops toggle and adds none, which contradicts both the policy and the sibling pattern on the same provider.

[github-actions]

Action items

• [medium] [violation] providers/ollama-cloud/models/deepseek-v4-flash:0731.toml:3 - Check: Every toggle must have a leading top-of-file wire-path comment. Why: reasoning_options includes { type = "toggle" } with no header documenting the Ollama request field (e.g. native think = true|false, or the OpenAI-compat path this provider exposes). Sync and consumers rely on that comment for the exact control surface. Action: Add a leading comment block above base_model with the exact toggle (and effort) wire syntax for ollama-cloud, matching AGENTS.md / the reasoning-options audit guide.

[renaudcerrato]

Addressed the toggle wire-path comment: added a leading top-of-file comment block documenting the native think = true|false toggle and the OpenAI-compat reasoning_effort/reasoning.effort effort path (high|max), sourced from the ollama-cloud provider.toml wire notes and https://docs.ollama.com/capabilities/thinking. bun validate still passes (exit 0).

[github-actions]

No actionable findings.