feat(providers): add Kyma

[sonpiaz]

Kyma is an OpenAI-compatible LLM gateway (https://api.kymaapi.com/v1) that gives one
endpoint and one key for open and frontier language models across many labs, plus speech,
embedding, rerank, image and video models. This adds providers/kyma/ with 60 models, each
base_model plus real overrides only, per AGENTS.md.

This supersedes #4019 (opened 2026-08-03, stale-closed 2026-09-03), which the automated
reviewer left two action items on. Both are fixed here:

1. reasoning_options was a blanket [] on 41 models. Every reasoning = true model here now
   carries real effort levels, sourced from that model's own lab hosting entry under
   models// where one exists, or from the effort set already established by other
   OpenAI-compatible relays in this repo for the same base_model. [] is kept only where the
   model is genuinely always-on with no caller control anywhere in the repo (MiniMax
   M2.5/M2.7, Kimi K2.7 Code, Grok Build, DeepSeek R1, Sonar/Sonar Pro).
2. Pricing is cited. Every [cost] value comes from https://api.kymaapi.com/v1/models, Kyma's
   public, unauthenticated model catalogue — also linked in a leading comment on each file.

60 of Kyma's 70 per-token chat models are included. 10 are left out because no models//
lab entry exists yet for the exact variant Kyma serves (very recent dated releases,
preview-only variants with no GA lab file, or one creator with no lab directory in this repo
yet) — happy to file those as a follow-up once there's a lab entry to point at, rather than
add stub lab files in the same PR as a new provider.

bun validate passes. logo.svg uses currentColor on a square viewBox, per AGENTS.md.

This entry is generated from the same catalogue that serves Kyma's own /v1/models, so prices
here and prices charged can't drift; happy to add a kyma:sync module in a follow-up if useful.

Maintained by Kyma — contact hello@kymaapi.com or @kyma-api if anything needs adjusting.

[github-actions]

Action items

• [high] [violation] providers/kyma/models/deepseek-v4-pro.toml:4 - Check: Relay reasoning_options must follow the lab + same-surface peer baseline for the underlying model. Why: DeepSeek V4 Pro lab and peers use toggle + effort high/max (lab maps low/medium→high). Kyma authors only effort high/max with no toggle, so callers cannot express the real on/off control this model family exposes. Same shape is wrong on deepseek-v4-flash.toml and deepseek-v4-flash-vision-exp.toml; flash lab also includes low. Action: For all three DeepSeek V4 models, author toggle + the lab effort set (high/max, plus low on Flash), with a leading wire comment for the Kyma request field that enables thinking.
• [high] [violation] providers/kyma/models/claude-haiku-4-5.toml:4 - Check: Do not invent a universal L/M/H effort set; copy lab/peer controls for this model. Why: Anthropic Haiku 4.5 is budget_tokens first-party; established relays use toggle (OpenRouter) or toggle + budget, not graded low/medium/high. Kyma invents GPT-style effort levels that neither the lab nor same-model peers use. Action: Replace with the Kyma-supported subset of lab/peer controls (typically toggle, and budget_tokens only if this host forwards a real reasoning budget), plus a leading toggle wire comment if toggle is kept.
• [high] [violation] providers/kyma/models/gemini-2.5-flash.toml:4 - Check: Gemini 2.5 Flash reasoning baseline is toggle/budget, not invented effort. Why: Google first-party is toggle + budget_tokens; OpenRouter peer is toggle. Kyma sets effort none/high, which is not the lab or peer surface for this model. Action: Align to toggle (and budget only if Kyma exposes a real reasoning budget), with a leading wire comment for toggle.
• [high] [violation] providers/kyma/models/qwen-3.6-plus.toml:4 - Check: Qwen hybrid reasoners on relays must not invent none/high in place of lab/peer toggle (and budget when applicable). Why: Alibaba first-party for Qwen3.6 Plus / 3.7 Plus / 3.7 Max / 3.7 Flash / 3-32b is toggle (+ budget_tokens); OpenRouter peers are toggle-based. Kyma flattens these to effort ["none","high"], inventing an off-via-effort control the baseline does not use. Same issue on qwen-3.7-plus.toml, qwen-3.7-max.toml, qwen3.7-flash.toml, qwen-3-32b.toml. Action: Replace with toggle (and budget only if this host forwards thinking_budget/equivalent); do not encode off as effort=none unless Kyma docs/tests show that wire behavior.
• [high] [violation] providers/kyma/models/qwen3.8-flash.toml:4 - Check: Qwen3.8 Flash lab baseline is toggle + graded effort + optional budget, not binary none/high. Why: Lab (providers/alibaba/models/qwen3.8-flash.toml) is toggle + low/medium/xhigh (+ budget). OpenRouter peer is toggle + budget_tokens. Kyma’s none/high matches neither. Action: Copy the lab/peer intersection Kyma actually forwards (at least toggle; effort and/or budget only if exposed), with a leading toggle wire comment.
• [high] [violation] providers/kyma/models/kimi-k2.5.toml:4 - Check: Kimi K2.x hybrid models use toggle, not invented none/high effort. Why: Moonshot lab and OpenRouter peers use toggle only for K2.5/K2.6. Kyma authors effort ["none","high"], which invents graded effort and mis-models off. Same on kimi-k2.6.toml. Action: Use { type = "toggle" } (with leading wire comment) unless Kyma documents a different real control set.
• [high] [violation] providers/kyma/models/kimi-k3.toml:4 - Check: Kimi K3 baseline is toggle + effort low/high/max. Why: Lab is explicit: thinking.type on/off plus adaptive effort low/high/max. Kyma keeps effort but drops toggle, so the separate on/off control is missing. Action: Author toggle + ["low","high","max"] with a leading toggle wire comment matching Kyma’s request syntax.
• [high] [violation] providers/kyma/models/glm-4.5-air.toml:4 - Check: GLM toggle-only models must not be rewritten as none/high effort. Why: Zhipu first-party uses toggle for GLM-4.5-Air, GLM-4.7-Flash, and GLM-5.1. Kyma invents effort ["none","high"] on all three (glm-4.7-flash.toml, glm-5.1.toml). Action: Use toggle (leading wire comment) unless Kyma documents a different real control.
• [high] [violation] providers/kyma/models/minimax-m3.toml:4 - Check: MiniMax-M3 lab baseline is toggle-only. Why: providers/minimax/models/MiniMax-M3.toml is { type = "toggle" }. Kyma invents effort ["none","high"]. Action: Match lab toggle (with wire comment) or document a Kyma-specific control that is actually exposed.
• [high] [violation] providers/kyma/models/mimo-v2.5.toml:4 - Check: MiMo-V2.5 lab baseline is toggle-only. Why: Xiaomi first-party is toggle; Kyma invents effort ["none","high"]. Action: Use toggle with a leading wire comment, or prove Kyma exposes a different control.
• [high] [violation] providers/kyma/models/gemma-4-31b.toml:4 - Check: Gemma 4 31B IT lab baseline is toggle-only. Why: providers/google/models/gemma-4-31b-it.toml is reasoning_options = [{ type = "toggle" }]. Kyma invents effort ["none","high"]. Action: Align to toggle (wire comment) unless Kyma has a verified different surface.
• [high] [violation] providers/kyma/models/claude-sonnet-5.toml:4 - Check: Claude Sonnet 5 lab includes a separate thinking toggle plus graded effort. Why: providers/anthropic/models/claude-sonnet-5.toml is toggle + low/medium/high/xhigh/max. Kyma keeps effort but omits toggle, understating the on/off control. Action: Add toggle alongside the effort list and document the Kyma wire path in a leading comment.
• [medium] [possible mistake] providers/kyma/models/qwen-3.8-max.toml:4 - Check: Qwen3.8 Max lab baseline is toggle + low/medium/xhigh (+ budget), not a full minimal→xhigh effort dump without toggle. Why: Lab is toggle + low/medium/xhigh + budget_tokens. Kyma matches OpenRouter’s broader minimal/low/medium/high/xhigh effort-only set, which drops toggle and adds levels beyond the lab. Action: Verify Kyma’s actual request surface; prefer lab intersection (toggle + real effort levels, budget only if present) over copying an expanded peer enum without host evidence.
• [medium] [possible mistake] providers/kyma/models/qwen3.8-27b.toml:4 - Check: Qwen3.8-27B peer baseline uses toggle + graded effort, not none inside effort. Why: OpenRouter peer is toggle + low/medium/xhigh. Kyma uses none/low/medium/xhigh with no toggle, which is a different off model than peers. Action: Confirm whether Kyma exposes toggle vs effort=none; align to the real wire control and add a toggle comment if toggle is used.

[sonpiaz]

Addressed every action item in the bot review; reasoning_options now copies the lab entry or the same-surface relay baseline per model instead of an invented effort set:

• Lab baseline (toggle + the lab's effort levels, budget dropped since Kyma has no numeric budget field): deepseek-v4-pro (high,max), deepseek-v4-flash and -vision-exp (low,high,max), kimi-k3 (low,high,max), claude-sonnet-5 (low,medium,high,xhigh,max), qwen-3.8-max (low,medium,xhigh).
• Lab baseline, toggle only: qwen-3.6-plus, qwen-3.7-plus, qwen-3.7-max, qwen3.7-flash, qwen-3-32b, kimi-k2.5, kimi-k2.6, glm-4.5-air, glm-4.7-flash, glm-5.1, minimax-m3, mimo-v2.5, gemma-4-31b.
• Same-surface relay baseline where the lab has no toggle or no entry: claude-haiku-4-5 (toggle), gemini-2.5-flash (toggle), qwen3.8-flash (toggle), qwen3.8-27b (toggle + low,medium,xhigh).

The wire comment on each file states the mechanism as verified today against Kyma's endpoint on qwen-3.6-plus: 200 reasoning tokens by default, 0 with reasoning_effort = "none". bun validate passes.

[github-actions]

Action items

• [high] [violation] providers/kyma/models/gemini-3.7-flash.toml:4 - Check: Reasoning options must match the lab + same-surface peer baseline, not invent levels. Why: Kyma sets effort to ["minimal", "low", "medium", "high"], but Google first-party and OpenRouter both use only ["low", "medium", "high"] for Gemini 3.7 Flash. minimal is not part of this model’s control set. Action: Change to [{ type = "effort", values = ["low", "medium", "high"] }] (same fix for gemini-3.8-flash.toml, which has the same mismatch).
• [high] [possible mistake] providers/kyma/models/claude-opus-4-7.toml:4 - Check: Multi-model relays should expose the lab/peer control set this host can actually forward. Why: Same-surface peers (e.g. OpenRouter) author toggle + effort for Claude Opus 4.7 / Opus 5 / Sonnet 4.6. Kyma already claims a live reasoning_effort = "none" off path and correctly used toggle (+ effort) on Claude Sonnet 5 and many other models, but these three stay effort-only with no toggle. Action: Either add { type = "toggle" } (with the wire comment) alongside the existing effort lists for claude-opus-4-7, claude-opus-5, and claude-sonnet-4-6 if Kyma honors none/off on those routes, or document host-specific evidence that off is not available for these Claude IDs.
• [medium] [possible mistake] providers/kyma/models/qwen3.8-flash.toml:4 - Check: Baseline effort from the lab entry for the same model. Why: Alibaba first-party qwen3.8-flash exposes toggle + effort low/medium/xhigh (+ budget). Kyma is toggle-only here, while the same PR already carries lab effort for qwen3.8-max. If Kyma forwards reasoning_effort, graded levels are missing. Action: Align with the lab effort set (toggle + ["low", "medium", "xhigh"]) unless Kyma-specific evidence shows only binary on/off on this ID; do not invent budget_tokens without a real budget field on this host.
• [medium] [possible mistake] providers/kyma/models/gemini-2.5-flash.toml:12 - Check: Provider limit.output overrides must be real host deltas, not placeholder caps. Why: Many entries force output = 8_192 (Gemini 2.5/3.x Flash family, gpt-5.6-terra, Grok 4.5/4.6, Muse Spark/Glimmer, Step 3.7 Flash, Gemma 4 31B, Hy3, GPT-OSS 120B, etc.) while lab metadata is typically 32k–500k. That pattern looks like a shared default rather than per-model catalogue values. Action: Confirm each limit.output (and the similar reduced contexts) against Kyma’s model catalogue/docs and keep only verified host caps; cite the catalogue field used.
• [medium] [possible mistake] providers/kyma/models/muse-spark-1.1.toml:14 - Check: [modalities] overrides must reflect this host, not expand lab capability. Why: Lab meta/muse-spark-1.1 is ["text", "image", "pdf", "video"]; Kyma adds audio, which is not on the lab entry. Other modality cuts (dropped pdf/video on Claude/Gemini/Kimi/Qwen/etc.) may be real host limits but are also uncited. Action: Remove audio unless Kyma documents audio input for Muse Spark 1.1; for other modality/limit overrides, keep only verified host deltas and note the source in the leading comment.

[sonpiaz]

Round two addressed in cdcf8a6, with host evidence measured on Kyma today (max_tokens 120, "What is 17*23? Think step by step."):

• claude-opus-5: 32 reasoning tokens by default, 0 with reasoning_effort = "none", so toggle is added next to its effort list.
• claude-opus-4-7 and claude-sonnet-4-6: 0 reasoning tokens by default on this host; thinking runs only when an effort is requested. Kept effort-only and wrote that measurement into the wire comment rather than claiming a toggle we did not observe.
• gemini-3.7-flash and gemini-3.8-flash: effort is now exactly low, medium, high.
• qwen3.8-flash: toggle + low, medium, xhigh per the lab entry; no budget field on this host.
• Limits: every [limit] override is removed; the entries inherit the lab context and output limits. The values Kyma's catalogue carried were defaults, not host caps, and Kyma imposes none of its own.
• Modalities: every [modalities] override is removed as well, so muse-spark-1.1 and the others inherit the lab set exactly.

bun validate passes.

[github-actions]

No actionable findings.

[sonpiaz]

Follow-up after a registry audit on our side. Three items:

qwen-3.6-plus — this id already inherits context = 1_000_000 / output = 65_536 from the alibaba/qwen3.6-plus lab entry (round 2 correctly dropped the stale 131_072/32_768 override), so no [limit] change is needed here. What was missing is modality scope: the lab entry lists video input, but this route only accepts text and image. Verified live today: a 532,316-token request returned 200 on this model with cost billed at the lab's >256K pricing tier (confirming the 1M context is real, not just inherited on paper), and a separate image-input request returned a correct answer. A video-input request is not part of this route's surface. Added [modalities] input = ["text", "image"] to reflect that real host delta, matching what our public /v1/models endpoint publishes for this id.

gpt-oss-120b and llama-3.3-70b — checked both against their lab entries (openai/gpt-oss-120b = reasoning = true, meta/llama-3.3-70b-instruct = reasoning = false) and both already resolve correctly via base_model inheritance with no override needed. No change on either file.

deepseek-v3 — this id actually serves DeepSeek V3.2 (hybrid thinking/non-thinking), not the original V3; base_model was pointing at the wrong lab entry (deepseek/deepseek-v3, non-reasoning) instead of deepseek/deepseek-v3.2 (reasoning). Re-pointed it so reasoning correctly resolves to true. We don't forward a caller-facing reasoning toggle on this route, so reasoning_options = [] (same shape as deepseek-r1 elsewhere in this PR) rather than inventing a toggle/effort set we haven't verified.

bun validate passes; bun test shows the same 6 pre-existing failures as a clean dev checkout (snapshot build artifact + sync-script tests, unrelated to any provider).

[github-actions]

Action items

• [medium] [violation] providers/kyma/models/nemotron-3-ultra-550b.toml:4 - Check: Relay reasoning_options must follow the lab + same-surface peer baseline for that model, not an invented control set. Why: Kyma authors effort none/medium/high. First-party providers/nvidia/models/nvidia/nemotron-3-ultra-550b-a55b.toml is toggle-only; OpenRouter peers use toggle + effort medium/high (plus budget on that surface). The Kyma set matches neither baseline and invents a none-in-effort shape the lab does not expose. Action: Replace with lab/peer-aligned options for this host (typically { type = "toggle" } or toggle + ["medium", "high"] if graded effort is actually forwarded); document the real wire path; do not keep the invented none/medium/high effort-only list.
• [medium] [possible mistake] providers/kyma/models/deepseek-v3.toml:8 - Check: On relays, reasoning_options = [] means affirmative no caller control, not uncertainty; hybrid DeepSeek V3.2 peers expose an on/off control. Why: base_model = "deepseek/deepseek-v3.2" is correct for a hybrid reasoner, but [] conflicts with established OpenAI-compatible peers (openrouter, novita-ai, tokengo, etc. use toggle) and with this PR’s own claim that Kyma forwards reasoning_effort = "none" on other hybrid routes. Action: Re-check this model id on Kyma; if off/on works, author toggle (with a model-accurate wire comment). Keep [] only with host-specific evidence that this route truly has no control.
• [low] [possible mistake] providers/kyma/models/claude-haiku-4-5.toml:2 (and the same copied toggle comment on the other non-qwen files that reuse it) - Check: Every toggle needs a leading wire comment for this route; do not claim live verification of a different model id. Why: Comments state “Verified live … on qwen-3.6-plus” on Claude, DeepSeek V4, GLM, Kimi, MiniMax, Gemma, etc. That overstates evidence for those routes and can hide real per-model differences (this PR already treats Claude Opus 4.7 / Sonnet 4.6 differently). Action: Rewrite toggle comments to the host wire field only, or cite verification for the same model id; drop cross-model “verified on qwen-3.6-plus” claims.