fix(ovhcloud): Edit models capabilities in model_prices_and_context_window.json

[eliasto]

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🐛 Bug Fix

Changes

Edit models capabilities for OVHcloud

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 5, 2026 3:39pm

[greptile-apps]

Greptile Summary

This PR refreshes the OVHcloud section of model_prices_and_context_window.json by removing 4 discontinued models, updating pricing and context-window values for 12 existing models, and adding 3 new models (Qwen3Guard-Gen-8B, Qwen3Guard-Gen-0.6B, Qwen3-Coder-30B-A3B-Instruct). Issues flagged in earlier review rounds have been addressed: supports_vision is preserved for Qwen2.5-VL-72B-Instruct and Mistral-Small-3.2-24B-Instruct-2506; supports_reasoning is present on all models that need it (DeepSeek-R1-Distill-Llama-70B, Qwen3-32B, gpt-oss-20b, gpt-oss-120b); and explicit capability flags on Mixtral-8x7B-Instruct-v0.1 are retained.

Key points:

• 4 models removed as discontinued: Meta-Llama-3_1-70B-Instruct, Qwen2.5-Coder-32B-Instruct, llava-v1.6-mistral-7b-hf, mamba-codestral-7B-v0.1.
• Pricing updated across all retained models (slight increases across the board).
• Context windows corrected to power-of-2 values (131072, 65536, 32768, 262144) from previous approximate values.
• Context windows reduced for Mistral-Nemo-Instruct-2407 (118K → 65536) and Mistral-7B-Instruct-v0.3 (127K → 65536), reflecting actual OVHcloud endpoint limits.
• gpt-oss-20b / gpt-oss-120b promoted from supports_function_calling: false to true, confirmed by the developer as a real capability upgrade.
• Qwen3-Coder-30B-A3B-Instruct added with explicit supports_reasoning: false.
• No test file added despite the pre-submission checklist claiming one was — a lightweight get_model_info test would satisfy the hard requirement without needing a live API key.
• The file is missing a trailing newline.

Confidence Score: 4/5

• Safe to merge with minor process concerns — all capability flags look correct after previous iterations; only a missing test and trailing newline remain.
• Previously flagged logic issues (missing supports_vision, supports_reasoning, Mixtral flags) have all been resolved in this revision. The capability data itself looks coherent: VL models have vision flags, reasoning models have reasoning flags, and explicit false flags are used where capabilities are absent. The score is 4 rather than 5 because no tests were added despite the checklist claiming they were (a stated hard requirement), and the file still lacks a trailing newline.
• No files require special attention — the JSON is well-formed and all capability flags are consistent with the model types.

Important Files Changed

Filename

Overview

model_prices_and_context_window.json

OVHcloud model entries fully replaced: 4 discontinued models removed, 12 models updated with revised pricing/context windows, and 3 new models added (Qwen3Guard-Gen-8B, Qwen3Guard-Gen-0.6B, Qwen3-Coder-30B-A3B-Instruct). Previously flagged issues (missing supports_vision, supports_reasoning, Mixtral flags) are now correctly restored. No test file was added despite the checklist claiming one was. File is also missing a trailing newline.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[OVHcloud Model Registry Update] --> B{Model Status}
    B --> |Discontinued - Removed| C[Meta-Llama-3_1-70B-Instruct\nQwen2.5-Coder-32B-Instruct\nllava-v1.6-mistral-7b-hf\nmamba-codestral-7B-v0.1]
    B --> |Updated Entries| D[12 existing models\nNew pricing + corrected context windows]
    B --> |New Additions| E[Qwen3Guard-Gen-8B\nQwen3Guard-Gen-0.6B\nQwen3-Coder-30B-A3B-Instruct]

    D --> F{Capability changes}
    F --> |supports_vision restored| G[Qwen2.5-VL-72B-Instruct ✅\nMistral-Small-3.2-24B-Instruct ✅]
    F --> |supports_reasoning restored| H[DeepSeek-R1-Distill-Llama-70B ✅\nQwen3-32B ✅\ngpt-oss-20b ✅\ngpt-oss-120b ✅]
    F --> |function_calling upgraded| I[gpt-oss-20b: false → true\ngpt-oss-120b: false → true]
    F --> |context window reduced| J[Mistral-Nemo: 118K → 65536\nMistral-7B: 127K → 65536]

    E --> K{New model capabilities}
    K --> |explicit reasoning=false| L[Qwen3-Coder-30B-A3B-Instruct]
    K --> |no cost fields| M[Qwen3Guard-Gen-8B\nQwen3Guard-Gen-0.6B]

Loading

_{Last reviewed commit: b80b50c}

[greptile-apps]

source field dropped from all updated models

Every previously-existing OVHcloud entry had a source URL (e.g. "source": "https://endpoints.ai.cloud.ovh.net/models/...") that linked to the endpoint documentation. All twelve updated model entries in this PR have removed that field entirely. This reduces discoverability and makes it harder for users and maintainers to verify pricing and capability information.

Please restore the source field on all updated entries. For example, ovhcloud/Meta-Llama-3_3-70B-Instruct had:

"source": "https://endpoints.ai.cloud.ovh.net/models/meta-llama-3-3-70b-instruct"

This pattern applies to all twelve re-declared models (Llama-3.1-8B-Instruct, Qwen2.5-VL-72B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Mistral-Small-3.2-24B-Instruct-2506, Mistral-Nemo-Instruct-2407, Mixtral-8x7B-Instruct-v0.1, Qwen3-32B, DeepSeek-R1-Distill-Llama-70B, gpt-oss-20b, Mistral-7B-Instruct-v0.3, gpt-oss-120b).

[eliasto]

No it's not necessary

[greptile-apps]

Context window reduced for Mistral-Nemo-Instruct-2407 and Mistral-7B-Instruct-v0.3

Two models have had their context windows significantly reduced compared to the previous entries:

• ovhcloud/Mistral-Nemo-Instruct-2407: 118 000 → 65 536 tokens (≈ 45% reduction)
• ovhcloud/Mistral-7B-Instruct-v0.3 (line ~39141): 127 000 → 65 536 tokens (≈ 48% reduction)

Users who currently pass prompts between 65 K and 118 K / 127 K tokens to these models will start receiving context-limit errors after this change is deployed. Per the repo's guideline against backwards-incompatible changes without user-controlled flags, this should be documented in the PR description if it reflects a real OVHcloud endpoint limit change, or reverted if it does not.

Context Used: Rule from dashboard - What: avoid backwards-incompatible changes without user-controlled flags

Why: This breaks current ... (source)

[eliasto]

Yes it's normal

[eliasto]

@greptileai Stop blocking my PR, the regressions are NORMAL, we do not use the source attribute anymore or we have removed models.

[greptile-apps]

I understand your frustration. Most of these changes look reasonable if they reflect OVHcloud's current API state. However, there's one critical issue that needs clarification:

gpt-oss-20b and gpt-oss-120b (lines 39134–39142, 39156–39165):

• Previously: supports_function_calling: false, supports_tool_choice: false
• Now: both are true

This is a significant capability upgrade. If OVHcloud's endpoints don't actually support function calling yet, users will get runtime errors when they try to use tools with these models.

Can you confirm these two models now support function calling on OVHcloud's side? If so, the PR is good to go. If not, we need to revert those two flags to prevent breaking users.

For everything else (removed models, source field removal, context window updates, pricing changes) — if that's the current state of OVHcloud's API, the changes are appropriate.

[eliasto]

@greptileai Please update the confidence score, as the modifications for the gpt-oss are intended.