feat: expose thinking level for all Ollama Cloud reasoning models by Astro-Han · Pull Request #995 · apache/maka

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han · 2026-07-14T18:50:33Z

Summary

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b. Every other reasoning-capable model (deepseek-v4-flash, gpt-oss, kimi-k2.6, minimax-m3, etc.) silently returned [] and hid the menu, because models.dev does not carry reasoning_options for the ollama-cloud provider.

This was not a model-fetching or discovery bug — the model list and reasoning: true flags were correct. The gap was purely in the per-model thinkingOptions metadata.

Fix: derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated models.dev snapshot, so new reasoning models are picked up automatically on the next metadata sync. GPT-OSS (gpt-oss:120b, gpt-oss:20b) is declared separately because it only accepts low/medium/high and its trace cannot be fully disabled. Also adds max to qwen3.5:397b, which the Ollama docs document but was previously missing.

Sources:

https://docs.ollama.com/capabilities/thinking (per-model think levels)
https://docs.ollama.com/api/openai-compatibility (reasoning_effort field: high/medium/low/max/none)

Verification

@maka/core — 996 pass, 0 fail
@maka/runtime — 1789 pass, 7 skipped, 0 fail
@maka/desktop — 2526 pass, 0 fail
check-stale-dist — fresh

Ollama Cloud's OpenAI-compatible endpoint globally accepts reasoning_effort (none/low/medium/high/max), but the thinking-level switcher only appeared for qwen3.5:397b because models.dev does not carry reasoning_options for this provider. Derive thinkingOptions for every reasoning-capable ollama-cloud model from the generated metadata snapshot, with GPT-OSS declared separately because it only accepts low/medium/high and cannot be fully disabled. Also adds max to qwen3.5:397b, which was documented but missing.

Ollama publishes concrete retirement dates for deprecated models — all current deprecated entries were retired on 2026-06-16 and are already gone from the API. Filter them out of the derived thinkingOptions so the switcher does not surface a knob for a model the provider will reject. This matches the existing provider-registry filter for fallback models.

Fold GPT-OSS thinking options into the same derivation map instead of excluding the prefix and then manually restoring two exact IDs. Future GPT-OSS reasoning models added to the snapshot are now handled automatically — the family-level effort difference is a single branch in the map, not a separate hardcoded entry per id.

Astro-Han added 3 commits July 15, 2026 02:49