provider: Add Nscale

[robputt]

Adds Nscale provider and models.

Pricing and model details: https://nscale.com/serverless
API docs: https://docs.nscale.com/docs
Base URL and auth: https://docs.nscale.com/docs/use-cases/chat

[github-actions]

Action items

• [high] [violation] providers/nscale/logo.svg:1 - Check: New provider logo must use currentColor with no fixed size and preferably a square viewBox. Why: The SVG declares width="157" height="89" (fixed size), a non-square viewBox="0 0 157 89", and fill="#111218" on all three paths instead of currentColor, so it will not adapt to light/dark themes and will not scale. A new provider without a compliant logo is a merge blocker. Action: Remove the width/height attributes, switch every fill="#111218" to fill="currentColor", and prefer a square viewBox (e.g. 0 0 24 24 with paths scaled accordingly).
• [high] [violation] providers/nscale/models/Qwen/Qwen3-235B-A22B.toml:1 - Check: Provider models must use base_model when a matching models/<provider>/<model>.toml entry exists. Why: models/alibaba/qwen3-235b-a22b.toml already defines the provider-agnostic facts for this model (name, description, reasoning, release_date, limit, modalities, open_weights, weights, benchmarks). This file re-declares all of them inline, duplicating catalog metadata. Action: Replace the inline definition with base_model = "alibaba/qwen3-235b-a22b" and keep only provider-specific overrides (reasoning_options, [cost], any [limit] override, and date overrides if needed).
• [medium] [possible mistake] providers/nscale/models/Qwen/Qwen3-235B-A22B-Instruct-2507.toml:3 - Check: reasoning/reasoning_options must match the served model variant. Why: The Qwen3-235B-A22B-Instruct-2507 is the non-thinking split of the Qwen3 235B family; established providers mark it reasoning = false with no reasoning_options (e.g. providers/togetherai/models/Qwen/Qwen3-235B-A22B-Instruct-2507-tput.toml, providers/nebius/models/Qwen/Qwen3-235B-A22B-Instruct-2507.toml). This file sets reasoning = true with a toggle, which contradicts the catalog convention and is unsupported by any citation. The same applies to Qwen3-4B-Instruct-2507.toml:3. Action: Verify against Nscale's docs which variant is served; if it is the Instruct (non-thinking) variant, set reasoning = false and remove reasoning_options (the schema rejects reasoning_options when reasoning = false).
• [medium] [violation] providers/nscale/models/Qwen/Qwen3-235B-A22B.toml:5 - Check: Niche-provider reasoning_options toggles must be documented with an adjacent TOML comment giving the exact API request syntax, and be backed by a citation. Why: Every reasoning_options = [{ type = "toggle" }] declaration (in Qwen3-235B-A22B.toml, Qwen3-4B-Thinking-2507.toml, and the three DeepSeek-R1-Distill-*.toml files) has no adjacent comment describing Nscale's toggle syntax and no PR citation showing Nscale exposes one. DeepSeek-R1-Distill and the Qwen3 "Thinking-2507" variant are always-reasoning models that conventionally expose no off toggle (existing distill entries use reasoning_options = []). Action: For each remaining reasoning model, either add an adjacent comment documenting Nscale's exact request syntax (e.g. # API: {"chat_template_kwargs": {"enable_thinking": false}}) plus a source citation, or use reasoning_options = [] when no verified toggle control exists.
• [medium] [possible mistake] providers/nscale/models/meta-llama/Llama-3.1-8B-Instruct.toml:9 - Check: knowledge (training cutoff) must not post-date release_date. Why: knowledge = "2024-12" is after release_date = "2024-07-23", which is impossible for a training-data cutoff. The same inconsistency appears in Llama-3.2-11B-Vision-Instruct.toml:9 (knowledge = "2024-12" vs release_date = "2024-09-25"). Meta's published cutoff for these families is December 2023. Action: Correct the knowledge values to a date on or before the release date (likely 2023-12), or cite a Nscale/Meta source for a different cutoff.
• [low] [violation] .pr-review/pull-request.json (PR body) - Check: Data-changing PRs should cite direct provider pricing, model documentation, or API references, with each citation stating what it supports. Why: The body only says "Adds Nscale provider and models." with no links, so pricing (e.g. DeepSeek distill 0.05–0.20, Mixtral 1.20, gpt-oss 0.05–0.40), context limits, capability flags, and the reasoning-toggle claims cannot be reviewed against a source. Action: Add links to Nscale's pricing/model docs and API reference, mapping each to the fields it justifies.

[github-actions]

Action items

• [high] [violation] providers/nscale/models/Qwen/Qwen3-235B-A22B-Instruct-2507.toml:1 - Check: New provider models must use base_model when a matching models/<provider>/<model>.toml exists, instead of duplicating provider-agnostic facts inline. Why: models/alibaba/qwen3-235b-a22b.toml exists for this underlying model, and several existing providers (Scaleway, Friendli, W&B) already serve the Instruct-2507 variant via base_model = "alibaba/qwen3-235b-a22b". W&B specifically demonstrates the correct pattern for the non-thinking variant: base_model plus a reasoning = false override. This file instead re-declares attachment, tool_call, structured_output, temperature, open_weights, modalities, and limit.context verbatim from the metadata entry. Action: Replace the inline definition with base_model = "alibaba/qwen3-235b-a22b", keep reasoning = false (with no reasoning_options), and retain only provider-specific overrides (name, description, release_date, last_updated, [cost], and the limit.input/limit.output overrides).
• [medium] [possible mistake] providers/nscale/models/openai/gpt-oss-120b.toml:3 (and providers/nscale/models/openai/gpt-oss-20b.toml:3) - Check: reasoning_options must reflect the reasoning controls the inference provider actually exposes. Why: Both files declare reasoning_options = [], but gpt-oss is built around the reasoning_effort parameter (low/medium/high) from the standard OpenAI Chat Completions API, and Nscale uses @ai-sdk/openai-compatible against an OpenAI-compatible endpoint. Five or more other providers serving these same models (Cerebras, Groq, Scaleway, OVHcloud, HuggingFace) expose reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }], several with first-party doc citations. An empty array is only correct when no caller control is verified, and the strong evidence here suggests control is available. Action: Verify against Nscale's API docs whether reasoning_effort is accepted for gpt-oss; if so, replace reasoning_options = [] with the effort option (and add a leading comment citing the Nscale doc that confirms the parameter).

[github-actions]

No actionable findings.