Docs preview: https://3febc4fb.dd-docs-preview.pages.dev

Notebook tutorials are placeholder-only in previews.

Greptile Summary

Updates the default nvidia-vision model alias from nvidia/nemotron-nano-12b-v2-vl to nvidia/nemotron-3-nano-omni-30b-a3b-reasoning and adjusts inference parameters (temperature lowered from 0.85 to 0.60). Changes are consistently reflected across the constants file, documentation, and test assertions.

Important Files Changed

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[PREDEFINED_PROVIDERS_MODEL_MAP\nnvidia / vision] --> B["model: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"]
    A --> C[inference_parameters:\nNEMOTRON_3_NANO_OMNI_30B_A3B_REASONING_INFERENCE_PARAMS]
    C --> D["temperature=0.60\ntop_p=0.95"]
    B --> E[nvidia-vision alias\nexposed to users]
    E --> F[test_default_model_settings.py\nassertion updated]
    E --> G[default-model-settings.md\ndoc table updated]
    E --> H[model-configs.md\nexample snippet updated]

_{Reviews (3): Last reviewed commit: "Merge branch 'main' into nmulepati/feat/..." | Re-trigger Greptile}

Review: PR #583 — feat: update default nvidia-vision model to nemotron-3-nano-omni-30b-a3b-reasoning

Summary

Small, focused config-only change that re-points the nvidia-vision alias in PREDEFINED_PROVIDERS_MODEL_MAP from nvidia/nemotron-nano-12b-v2-vl to nvidia/nemotron-3-nano-omni-30b-a3b-reasoning, tuned with temperature=0.60, top_p=0.95. A new NEMOTRON_3_NANO_OMNI_30B_A3B_REASONING_INFERENCE_PARAMS constant is introduced, and docs + one test assertion are updated to match. 4 files, +8/-4.

Findings

Correctness

• Only the nvidia provider entry is updated; openrouter-vision is intentionally left on the old model. Confirmed via constants.py:378 in the PR branch — openrouter still references nvidia/nemotron-nano-12b-v2-vl with DEFAULT_VISION_INFERENCE_PARAMS. That is why DEFAULT_VISION_INFERENCE_PARAMS remains (correctly) defined. The test builtin_model_configs[10] at line 90 of test_default_model_settings.py still asserts the old name, which is consistent. Worth confirming in the PR description that the openrouter asymmetry is deliberate — the PR title and summary say "update default nvidia-vision model" which reads cleanly, but a reader skimming the diff may wonder whether the openrouter entry was overlooked.
• New model name advertises "reasoning" but no reasoning_effort is set. Every other "reasoning" model in this file (nemotron-3-super-120b-a12b, gpt-5) passes extra_body={"reasoning_effort": "medium"}. If nemotron-3-nano-omni-30b-a3b-reasoning supports the same knob, omitting it is a silent behavioral divergence; if it doesn't, a one-line comment here would pre-empt the obvious reviewer question. Please confirm with the model card.

Test coverage

• test_get_builtin_model_configs asserts the new model string but does not assert the new inference parameters (temperature=0.60, top_p=0.95). The inference params are half of what this PR changes — an accidental reversion to the old values would pass tests today. Recommend adding assert builtin_model_configs[2].inference_parameters.temperature == 0.60 and top_p == 0.95 (or asserting the full dict against the new constant) alongside the existing model-name assertion. Low-cost, high-signal.

Conventions & style

• Naming NEMOTRON_3_NANO_OMNI_30B_A3B_REASONING_INFERENCE_PARAMS matches the existing NEMOTRON_3_NANO_30B_A3B_INFERENCE_PARAMS / NEMOTRON_3_SUPER_120B_A12B_INFERENCE_PARAMS scheme — consistent.
• The placement of the new constant next to the other model-specific constant blocks is correct.
• No changes needed to imports, types, or __future__ annotations. No ruff-relevant surface touched.

Docs

• default-model-settings.md row for nvidia-vision is updated. Good.
• model-configs.md swaps the model string in the example but leaves the example's temperature=0.7, top_p=0.95, max_tokens=2048 unchanged. That's fine — it's an example of custom config, not a canonical default, and the values there have always differed from the real default. No action needed, just flagging for awareness.
• Stale reference: docs/devnotes/posts/assets/data-designer-got-skills/trace-skill.html:169 still shows nvidia-vision nvidia/nemotron-nano-12b-v2-vl. This looks like a captured trace artifact from a historical run, so it's reasonable to leave it as-is (trace outputs rot naturally), but worth a one-line mention in the PR if you want to be explicit that it was seen and intentionally skipped.

Performance / Security

• Nothing to flag. Pure config change; no new imports, no network code, no secrets surface.

Verdict

Approve with two small asks (both optional, neither blocking):

1. Extend test_get_builtin_model_configs to also assert the new temperature/top_p values so the inference-params half of this change is covered.
2. In the PR description (or a brief inline comment), confirm the reasoning_effort omission is intentional given the new model's "reasoning" suffix, and note that openrouter-vision is deliberately left on the previous model.

Change itself is low-risk, well-scoped, and follows established patterns in constants.py.