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

[airtaxi]

Summary

• adds Ollama Cloud deepseek-v4-flash:0731-cloud to the ollama-cloud provider catalog
• self-contained model definition mirroring the existing deepseek-v4-flash.toml entry, with :0731-cloud as the tagged revision ID
• release_date / last_updated set to 2026-07-31 matching the 0731 post-training revision
• advertises native high / max effort controls (plus toggle) for the OpenAI-compatible endpoint, consistent with sibling deepseek-v4-flash.toml

Sources

• https://ollama.com/library/deepseek-v4-flash

Verification

• bun validate

[github-actions]

Action items

• [high] [violation] providers/ollama-cloud/models/deepseek-v4-flash:0731-cloud.toml:1 - Check: Non-lab hosts must use base_model with an override-only provider file. Why: Ollama Cloud did not create this model; lab metadata already exists at models/deepseek/deepseek-v4-flash-0731.toml, and other ollama-cloud entries (e.g. kimi-k3, glm-5.2) correctly inherit. This file fully inlines lab facts instead of linking them. Action: Set base_model = "deepseek/deepseek-v4-flash-0731" and keep only provider deltas (name, reasoning_options, and real [limit] overrides such as context/output 1048576 if intentional). Drop restated identical fields (description, family, attachment, reasoning, tool_call, open_weights, dates, [modalities], etc.).
• [medium] [violation] providers/ollama-cloud/models/deepseek-v4-flash:0731-cloud.toml:6 - Check: Every toggle requires 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 vs OpenAI-compatible reasoning_effort / reasoning.effort per providers/ollama-cloud/provider.toml). Action: Add a leading comment block above the first key with the exact toggle (and effort) wire syntax this host exposes for the model.

[github-actions]

Action items

• [medium] [violation] providers/ollama-cloud/models/deepseek-v4-flash:0731-cloud.toml:1 - Check: Toggle vs effort=none wire path for reasoning_options. Why: The leading comment says OpenAI-compat off is reasoning_effort/reasoning.effort = "none", but the file still authors { type = "toggle" } with effort ["high", "max"] only. Per AGENTS.md, when off is an effort value that includes none, use effort alone (no toggle); toggle is only for a separate on/off control and needs that exact wire path documented. This host’s provider.toml documents the OpenAI /v1 surface as high|medium|low|max|none, and npm is @ai-sdk/openai-compatible against that API. Action: Pick one coherent surface: either drop toggle and set reasoning_options = [{ type = "effort", values = ["none", "high", "max"] }] (add low/medium only if this host actually exposes them for DeepSeek V4), or keep toggle only if a distinct field (e.g. native think=true|false) is the intended control and rewrite the top-of-file comment to that exact wire path—do not call none a toggle.

[github-actions]

No actionable findings.

[renaudcerrato]

Hey @airtaxi — opened #3985 as a corrected alternative covering the same model. Wanted to flag the differences here so they're in one place; happy to defer to maintainers on which to land.

The three corrections, with sources:

1. Model ID / filename. The Ollama Cloud API exposes deepseek-v4-flash:0731 (both /api/tags and /v1/models), with no -cloud/:cloud suffix. This PR uses that exact ID. The suffixed form here (:0731-cloud) has caused 400/404s in downstream consumers before — see NousResearch/hermes-agent#16179.

2. name. Per the override-only rule in AGENTS.md, name shouldn't be restated unless it truly differs; if it does, it shouldn't be the raw ID. feat(ollama-cloud): add deepseek-v4-flash:0731 model #3985 inherits the lab's display name "DeepSeek V4 Flash 0731" from base_model = "deepseek/deepseek-v4-flash-0731".

3. reasoning_options. AGENTS.md specifies toggle + high/max for DeepSeek V4 (no none). The existing ollama-cloud siblings — deepseek-v4-flash.toml and deepseek-v4-pro.toml — both use [{ toggle }, { effort: ["high","max"] }]. This PR matches that. The effort: ["none","high","max"] (no toggle) shape here drops the toggle and adds none, which contradicts both the policy and the sibling pattern.

Happy to collaborate if you'd rather fold these fixes into this PR instead.

[airtaxi]

Thanks for the review @renaudcerrato — confirmed. This PR is superseded by #3985 which covers the same model with the corrected Ollama Cloud ID (deepseek-v4-flash:0731) and the proper override-only structure. Closing this one in favor of #3985.