Add Arka-AI provider (OpenAI-compatible model hub)

[mitek]

Summary

Adds Arka-AI (https://arka-ai.ru) — an OpenAI-compatible model hub serving https://api.arka-ai.ru/v1 — with 23 chat/reasoning models from DeepSeek, Qwen (Alibaba), Kimi (Moonshot), MiniMax, GLM (Zhipu), and Step.

• Provider: name = "Arka-AI", npm = "@ai-sdk/openai-compatible", env = ["ARKA_API_KEY"], api = "https://api.arka-ai.ru/v1".
• logo.svg is the official Arka AI mark (single currentColor, no fixed size).

Models (23)

All provider files are override-only and base_model existing lab entries:

Arka model id	base_model
DeepSeek-V4-Pro	deepseek/deepseek-v4-pro
deepseek-v4-flash-0731	deepseek/deepseek-v4-flash-0731
DeepSeek-V3	deepseek/deepseek-v3
DeepSeek-R1	deepseek/deepseek-r1
GLM-5.2 / GLM-5.1 / GLM-5 / GLM-4.7-Flash	zhipuai/*
Qwen3.5-27B / -Plus / -Flash, Qwen3.6-Plus / -Max, Qwen3.7-Max / -Plus, qwen3.8-max, Qwen3-Coder-Next / -Flash	alibaba/*
MiniMax-M2.5 / MiniMax-M3	minimax/*
Kimi-K2.7-Code / kimi-k3	moonshotai/*
Step-3.7-Flash	stepfun/*

Notes

• reasoning_options are copied from each underlying lab's first-party provider entry (e.g. DeepSeek V4 = toggle + effort[high,max], Qwen = toggle + budget_tokens, Kimi K3 = toggle + effort[low,high,max]).
• interleaved = { field = "reasoning_content" } where the underlying host exposes it.
• Cost is converted from Arka AI's RUB list price at 82.9977 RUB/USD (CBR rate captured 2026-08-13). Sources: https://arka-ai.ru (pricing) and https://api.arka-ai.ru/v1/models.

bun validate passes.

[github-actions]

Action items

• [high] [violation] providers/arka-ai/models/DeepSeek-V4-Pro.toml:127 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: Arka-AI is a multi-model relay; 15 model files declare { type = "toggle" } but only have cost/source headers—no exact request field (e.g. thinking.type, enable_thinking, reasoning.enabled). Without that, clients cannot know how to turn reasoning on/off on this host. Action: Add a leading # Toggle: … (and matching effort/budget wire notes) to every file that includes toggle, documenting Arka’s actual request path for that model.
• [medium] [violation] providers/arka-ai/models/deepseek-v4-flash-0731.toml:451 - Check: Relay reasoning_options must match the lab first-party baseline for that model. Why: First-party providers/deepseek/models/deepseek-v4-flash.toml uses toggle + effort ["low", "high", "max"] (Flash keeps low; Pro maps low→high). This entry copies Pro’s ["high", "max"] only, so callers lose a real lab control level the PR claims to mirror. Action: Set Flash effort to ["low", "high", "max"] (or document host-specific evidence if Arka truly omits low).
• [medium] [possible mistake] providers/arka-ai/models/GLM-4.7-Flash.toml:147 - Check: Toggle must reflect this host’s controls, baseline lab/peers. Why: There is no first-party providers/zhipuai entry for glm-4.7-flash; peers disagree (openrouter uses [], vercel uses toggle). Authoring bare toggle (and no cost) without Arka/API evidence may invent a control this hub does not expose. Action: Confirm Arka’s GLM-4.7-Flash request surface and either keep toggle with a wire comment + evidence, or set reasoning_options = [] if thinking is always-on/uncontrolled here.

[mitek]

Addressed all three review items in 7001370, verified live against https://api.arka-ai.ru/v1/chat/completions (2026-08-18):

[high] Toggle wire-path comments — added top-of-file wire notes to all 15 toggle files. Live probing showed enable_thinking true|false is the canonical toggle on Arka (with thinking.type = enabled|disabled also accepted), except Qwen3.5-27B where only enable_thinking takes effect — exactly the per-model ambiguity the missing comments hid.

[medium] deepseek-v4-flash-0731 effort — now ["low", "high", "max"]; all three levels verified live on Arka, matching the first-party deepseek baseline (low is not mapped away here, unlike the alibaba-hosted copy).

[medium] GLM-4.7-Flash — keeping toggle with evidence: the first-party providers/zhipuai/models/glm-4.7-flash.toml entry does declare the toggle, and Arka's GLM siblings (GLM-5, GLM-5.1) were verified accepting both enable_thinking and thinking.type. The GLM-4.7-Flash route itself returned upstream_error (HTTP 400) for every request at verification time, noted in the file.

Two additional corrections surfaced by probing:

• kimi-k3: reasoning_options is now [] — disabling reasoning 400s ("invalid temperature: only 0.6 is allowed" from the non-thinking lane) and effort controls show no observable effect, so the host exposes no client control here.
• budget_tokens removed from 8 Qwen files: thinking_budget is accepted but not enforced (reasoning length unchanged at budget=64 vs 16384/32768, both streaming and non-streaming).

bun ./packages/core/script/validate.ts passes locally.

[github-actions]

Action items

• [medium] [possible mistake] providers/arka-ai/models/kimi-k3.toml:1 - Check: reasoning_options = [] only when this host truly has no caller control. Why: The header claims reasoning is always on, but the same note says enable_thinking=false / thinking.type=disabled hits the non-thinking lane with invalid temperature: only 0.6 is allowed. That implies an off path may work when temperature is set correctly, so [] may under-report controls versus the moonshot baseline (toggle + effort low|high|max). Action: Re-test disable with temperature=0.6 (and adaptive/output_config.effort if applicable). If off or effort works, restore the matching reasoning_options and a leading wire-path comment; if both still fail with correct params, keep [] and clarify that the non-thinking lane is unusable.

[github-actions]

No actionable findings.