adds FPT AI Factory as provider

[littlebeanhp]

Adds FPT AI Factory as an OpenAI-compatible provider (https://mkp-api.fptcloud.com/v1), serving 8 LLM/VLM models on NVIDIA GB300 nodes. Claude and GPT models are coming in an upcoming release.

Sources

All values are captured from FPT's live first-party catalog endpoint (public, no auth): GET https://mkp-api.fptcloud.com/v1/models.

• Model IDs — data[].id (exact-cased, e.g. gpt-oss-120b, DeepSeek-V4-Flash).
• [cost] — data[].pricing.prompt → input, data[].pricing.completion → output (per-token USD ×10⁶ = per-1M).
• [limit].context — data[].top_provider.context_length (the cap FPT actually serves; several are lower than the lab base, e.g. Llama-3.3-70B 128k→32k, DeepSeek-V4-Flash 1M→200k).
• [limit].output — FPT reports top_provider.max_completion_tokens == context_length for every model, which is a gateway "no separate output cap" signal rather than a real max-output. Following peer relays (tinfoil, aki-io, nearai), output is left to inherit the lab base_model's true max-output rather than copy the context window. Overrides are kept only where FPT genuinely differs: GLM-5.2 (128k output, below the 1M context and below base) and DeepSeek-V4-Flash (output bounded by its 200k FPT context).

Provider metadata: API keys at https://marketplace.fptcloud.com/ · docs at https://ai-docs.fptcloud.com/.

Reasoning options

reasoning_options = [] is intentional and evidence-backed: FPT's data[].supported_parameters for every served model is exactly ["temperature","top_p","max_tokens","tools","tool_choice","response_format","stream"] — it exposes no reasoning, reasoning_effort, thinking, or budget_tokens parameter. The gateway forwards no caller-facing reasoning control, so [] is affirmative (not "untested default").

Changes since first review

• Removed GLM-5.1 — FPT's live catalog no longer lists it (only GLM-5.2 is served).
• Corrected limit.output on the 6 models that had copied the context window; they now inherit the lab base max-output.
• Provider logo.svg is a square (viewBox ~1:1) icon mark using currentColor, no fixed size or hardcoded colors.

Validated with bun run validate.

[github-actions]

Action items

• [critical] [violation] providers/fpt-ai-factory/provider.toml:1 - Check: New providers must ship a compliant providers/<id>/logo.svg. Why: This PR adds provider fpt-ai-factory with no logo.svg. Missing logos are a hard merge blocker. Action: Add providers/fpt-ai-factory/logo.svg using currentColor, no fixed size or hardcoded colors, preferably a square viewBox.
• [medium] [possible mistake] providers/fpt-ai-factory/models/gpt-oss-120b.toml:10 - Check: limit.output should be the provider max completion tokens, not a copy of context. Why: Eight of nine models set output equal to context (e.g. gpt-oss 128_000, gemma 262_000, Qwen 262_144, Llama 32_000), while base metadata uses much smaller outputs (32_768, 65_536, 4_096). That pattern usually means max-output was filled with the context window. Action: Verify each model’s max output from FPT docs/API and correct limit.output (and any wrong limit.context) accordingly.
• [low] [possible mistake] providers/fpt-ai-factory/models/DeepSeek-V4-Flash.toml:5 - Check: Data-changing PRs should cite first-party pricing/limit sources. Why: All nine models introduce costs and limit overrides with no PR-body mapping to FPT pricing or model docs, so the figures cannot be reviewed. Action: Add direct FPT pricing/docs links in the PR body and state which cost and limit fields each source supports.

[github-actions]

Action items

• [medium] [possible mistake] providers/fpt-ai-factory/logo.svg:1 - Check: New-provider logo should work as a small square mark (currentColor, no fixed size/hardcoded color; prefer square viewBox). Why: The SVG is a full “FPT AI FACTORY” wordmark with viewBox="0 0 1536.61 222.32" (~7:1). The site renders provider logos in a 1rem × 1rem box, so this will scale down to a thin unreadable smear. Action: Replace with the icon/mark only and a square viewBox (e.g. cropped around the left glyph), still using currentColor and no fixed width/height or hardcoded colors.
• [low] [possible mistake] providers/fpt-ai-factory/models/DeepSeek-V4-Flash.toml:5 - Check: Data-changing PRs should cite first-party pricing/docs/API sources mapped to claims. Why: All nine models introduce provider-specific cost and limit values (several set output equal to context, and several diverge sharply from the linked base_model limits, e.g. Qwen/Gemma/gpt-oss max output), but the PR body only describes the provider and validation—no pricing page, model catalog, or API reference. Action: Add PR-body citations (what each URL supports) for model IDs, USD pricing, and context/output limits; adjust any values that the sources do not support.

[github-actions]

Action items

• [medium] [possible mistake] providers/fpt-ai-factory/models/gpt-oss-120b.toml:9 - Check: Provider-specific limit values should reflect this provider’s real context/max-output caps, not placeholders. Why: Nearly every model sets limit.output equal to limit.context, often far above the inherited base max output (e.g. openai/gpt-oss-120b 32_768→128_000; google/gemma-4-31b-it 32_768→262_000; meta/llama-3.3-70b-instruct 4_096→32_000; alibaba/qwen3.6-27b 65_536→262_144). Several also shrink context without evidence (DeepSeek-V4-Flash 1M→200k, Llama-3.3-70B-Instruct 128k→32k, GLM-5.1 200k→128k). That pattern usually means guessed caps rather than documented FPT limits. Action: Verify each model’s context and max output against FPT model/API docs; correct the values, or remove [limit] overrides so accurate base limits are inherited where FPT does not differ.
• [low] [possible mistake] providers/fpt-ai-factory/provider.toml:1 - Check: Data-changing PRs should cite first-party sources mapped to claims. Why: The PR introduces a full provider catalog (all costs plus custom limits) with only a generic docs homepage and no PR-body citations that state what supports each price or limit. Those values cannot be reviewed from the diff alone. Action: Add PR citations (pricing page, model list, or API reference) and note which claim each URL supports—costs and any non-inherited limits at minimum.

[github-actions]

Action items

• [high] [violation] providers/fpt-ai-factory/models/DeepSeek-V4-Flash.toml:2 - Check: On multi-model relays, reasoning_options = [] means affirmative no caller control, not uncertainty; baseline is lab + same-surface peers. Why: FPT is an OpenAI-compatible multi-lab gateway. DeepSeek first-party uses toggle + effort high/max, and established relays expose controls. The PR only cites the models catalog for pricing/limits, so empty options look like an untested default. Action: Replace [] with lab/peer-aligned options (typically toggle + ["high","max"] with a leading wire comment), or keep [] only with host docs/live proof that FPT exposes no reasoning control.
• [high] [violation] providers/fpt-ai-factory/models/GLM-5.1.toml:2 - Check: Relay reasoning_options must follow lab/peer controls for the model. Why: Lab/first-party Zhipu/Z.ai entries use toggle for GLM-5.1; empty options on this relay are not justified by the PR evidence. Action: Set reasoning_options to match the host (baseline toggle + leading wire comment), or document that FPT has no on/off control.
• [high] [violation] providers/fpt-ai-factory/models/GLM-5.2.toml:2 - Check: Relay reasoning baseline for GLM-5.2. Why: Zhipu/Z.ai first-party use effort high/max (not empty). Action: Author that effort set if FPT accepts reasoning_effort, or prove no control before keeping [].
• [high] [violation] providers/fpt-ai-factory/models/Qwen3.6-27B.toml:2 - Check: Relay reasoning baseline for Qwen3.6-27B. Why: Alibaba first-party uses toggle + budget_tokens; OpenRouter peer uses toggle. Empty options conflict with that baseline unless FPT truly has no control. Action: Prefer peer-aligned toggle (add budget_tokens only if this host has a real thinking budget field); do not leave [] from uncertainty.
• [high] [violation] providers/fpt-ai-factory/models/gemma-4-26B-A4B-it.toml:2 - Check: Relay reasoning baseline for Gemma 4 IT. Why: Google first-party and peers (e.g. OpenRouter) use toggle for these reasoners. Action: Set toggle with a leading wire comment if FPT forwards on/off, or evidence no control before []. Apply the same fix to gemma-4-31B-it.toml.
• [high] [violation] providers/fpt-ai-factory/models/gpt-oss-120b.toml:2 - Check: Relay reasoning baseline for GPT-OSS. Why: Established same-surface peers (OpenRouter, Groq, etc.) use effort low/medium/high for gpt-oss-120b/20b; empty options are not the peer baseline. Action: Set reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }] if FPT accepts reasoning_effort, or prove no control. Apply the same fix to gpt-oss-20b.toml.