Model providers and compatibility

Current availability

Known integration bugs; temporarily unavailable: MiMo Token Plan, Kimi API, Kimi Code, DeepSeek, Alibaba Token Plan, Volcano Coding Plan, and Volcano Ark API. Setup displays these entries in gray and prevents selection. Existing profiles are retained for viewing or removal; activation, model tests, refresh and execution are blocked before network access. No automatic backup engine or billing-route switch is used.

These entries are currently unsupported. Endpoint and protocol details below remain as historical reference, not instructions to enable an unavailable integration. This status overrides earlier compatibility claims. Other entries, including local endpoints, MiMo pay-as-you-go, and Alibaba Coding Plan, keep their separate status.

This guide covers 0.4.0, checked on October 23, 2026. Koder directly handles Chat Completions, Anthropic Messages, and Responses. It does not require a separate agent engine to adapt a model API.

What the catalog means

The 36 visible presets contain 108 curated model choices. Each cloud preset has a local native tool-call/result round trip. Local fixtures check request fields, response parsing and continuation; they do not establish your account’s model access, region, quota, billing or reliability. Some providers require model activation or a deployment endpoint in their console. The list contains task models, not every audio, image, embedding or batch service a vendor sells.

A model’s presence in a public listing does not guarantee access with every account. NVIDIA, Hugging Face and OpenRouter entries were compared with their live public model APIs. Other entries use the linked official model documentation. SiliconFlow’s current available models are account-specific; its official documented model example is bundled, and the account refresh below returns the exact IDs your key can access. Do not infer a reseller ID by copying the original vendor’s ID.

Tencent HY4 TokenHub is a separate endpoint and credential choice. It does not move legacy Hunyuan keys to a different host. Xiaomi pay-as-you-go and Token Plan keys also stay separate; see MIMO.md.

Verify your selected account

First save a provider in koder setup. Then run:

koder model test
koder model test --tools

The first command checks a text reply only. The second explicitly makes up to two model requests using a harmless fixed probe function and its result; the provider may bill those requests. It does not read files, run a shell command or grant computer permissions. Only successful tool calling and continuation produces the task-capability success message. A 401, permission failure, quota error or text-only result is reported as a failed check, without changing the account or selecting another model.

Get current account model IDs

koder model models
koder model models chatgpt
koder model refresh

The first two commands show the bundled catalog without network access. Refresh performs authenticated, read-only model-list requests to the configured endpoint. It writes a local metadata cache, without changing the active model, endpoint or API key and without making an inference request. Models explicitly marked as supporting tools can appear in the next setup picker. Entries with unknown tool support remain listed in the refresh output as unverified; enter an ID manually and run the tool check before relying on it. Unsupported model-list APIs fail clearly and keep the previous cache.

Removed or excluded choices

• Yi is hidden because its inherited provider notice marks the API as stopped. A saved official Yi endpoint receives a clear local error. Koder never moves that key to another service.
• Legacy Perplexity Sonar choices are removed after the provider’s announced September 27, 2026 retirement. Perplexity now uses the Agent API through its Responses alias and supports local custom function loops.
• Batch-only IDs, transcription/TTS, safety classifiers and dedicated computer-use API models are excluded from the general task picker. A third-party SDK catalog entry alone is not proof of task compatibility.
• Baichuan medical M3 variants are not offered as general task models until their model-specific tool contract is established. The documented Baichuan4 tool-capable models remain available.

This is a statement about this release’s supported task contracts, not a claim that every excluded model is impossible to integrate forever. Custom endpoints remain explicit user configurations. No account acceptance is claimed for any vendor in this repair.

Backup engine state

Koder is the default and handles the configured API directly. The internal fallback integration is retained but disabled and gray in setup by default. A saved old Hermes selection cannot launch it unless the internal fallback-enabled setting is already explicit. There is no automatic fallback after an API failure. The pinned fallback runtime cannot preserve a non-OpenAI Responses custom route and rejects that route before handing off a credential. Improve or change the native API configuration instead.

Download completion deadline

The installer bootstrap, program/runtime downloads and CLI updater default to 100000000 seconds. You may explicitly set KODER_DOWNLOAD_TIMEOUT_SECONDS to an integer from 1 through 100000000 for a shorter limit. The long total ceiling is separate from the no-progress guards: connection/header inactivity defaults to 30 seconds and body inactivity to 60 seconds. Set KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS (1–3600) and KODER_DOWNLOAD_IDLE_TIMEOUT_SECONDS (1–86400) explicitly when needed. Progress remains visible and Ctrl+C cancels. Long Node deadlines are scheduled in bounded chunks, avoiding timer overflow. This setting does not increase the five-second existing-Node probe or model API timeouts. Checksums and agreement checks still apply.

Alibaba and Volcano subscriptions

Checked October 20, 2026. These are separate native API routes; subscription access is not implied by an ordinary provider key. A custom proxy remains explicit and is never silently redirected.

Alibaba Cloud: aliyun-token-plan

Base URL: https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1. Default: qwen3.8-max. Models: qwen3.8-max, qwen3.8-flash, qwen3.7-max, qwen3.7-plus, qwen3.6-flash, deepseek-v4.1-flash, deepseek-v4.1-flash-1017, deepseek-v4-pro, deepseek-v4-pro-0813, deepseek-v4-flash-0731, glm-5.3, glm-5.2, auto.

Official setup. Account and regional availability must be checked in the plan console.

Alibaba Cloud: aliyun-coding-plan

Base URL: https://coding.dashscope.aliyuncs.com/v1. Default: qwen3.7-plus. Models: qwen3.7-plus, qwen3.6-plus, kimi-k2.5, glm-5, MiniMax-M2.5, qwen3-coder-next, qwen3-coder-plus.

Official setup. Account and regional availability must be checked in the plan console.

Volcano Engine: volcano-coding-plan

Base URL: https://ark.cn-beijing.volces.com/api/coding/v3. Default: ark-code-latest. Models: ark-code-latest, doubao-seed-2.1-pro, doubao-seed-2.1-lite, doubao-seed-evolving, doubao-seed-2.0-mini, glm-5.3, glm-5.3-flash, deepseek-v4.1-flash, deepseek-v4.1-flash-economy, deepseek-v4-flash, deepseek-v4-pro, kimi-k3, kimi-k2.8-preview.

Official setup. Account and regional availability must be checked in the plan console.

Alibaba Token Plan personal/team subscriptions share the documented Token Plan service path but have their own issued keys and data terms. Coding Plan is a separate subscription with different supported versions and no Responses route. Both use their own sk-sp- key; the prefix alone cannot prove which subscription issued it. Ordinary Alibaba API credentials do not substitute for a subscription.

Volcano Coding Plan uses /api/coding/v3 for Chat Completions or Responses, and /api/coding for Anthropic Messages. Standard /api/v3 is a separate metered route. Use ark-code-latest to follow the console selection, or an exact supported model name; metered ep- endpoint IDs are not Coding Plan model names.

Token Plan Anthropic base: https://token-plan.cn-beijing.maas.aliyuncs.com/apps/anthropic. Alibaba Coding Plan Anthropic base: https://coding.dashscope.aliyuncs.com/apps/anthropic. Koder appends /v1/messages for these documented compatibility bases.

Provider subscription rules and data practices remain the provider’s terms. In particular, review the current Alibaba personal/team data terms before submitting private work. No compatibility test here establishes live account access or billing eligibility.

Kimi, Kimi Code and DeepSeek recovery

Moonshot API supports Chat Completions and Responses at the official api.moonshot.cn/v1 and api.moonshot.ai/v1 bases. Its Anthropic compatibility base is /anthropic. DeepSeek also supports Responses; its Anthropic compatibility base is /anthropic. Use the protocol offered for your selected product. Kimi K3 uses max_completion_tokens and low/high/max reasoning effort; it does not offer a thinking-off setting. DeepSeek uses deepseek-flash or deepseek-v4-pro on api.deepseek.com. Deprecated official deepseek-chat/reasoner IDs fail with a specific local explanation instead of silently selecting a billed replacement. Custom proxy routes remain explicit. Official native and subscription routes require HTTPS. Bare official origins are normalized to the selected protocol; mixed protocol paths and cross-product paths fail locally. Model redirects fail safely rather than forwarding your key or task to another endpoint.

Kimi Code is a separate membership product with separate credentials: choose kimi-code in setup. Its default is k3-256k at https://api.kimi.com/coding/v1 using Responses. Overseas accounts may use https://api.kimi.ai/coding/v1. Select only an entitled model ID from the four listed choices. Unknown IDs are rejected to avoid a documented silent fallback. Koder identifies itself honestly as KoderCLI; membership or client restrictions are not bypassed. Context and model access depend on your tier, and a 401/403 can indicate an entitlement issue. K3 and K2.8 support low/high/max; HighSpeed keeps thinking on. No automatic route or model switch occurs.

For these documented native models, the default output reserve is 32768 tokens and reasoning starts low when supported. An explicit user context/output ceiling is retained. A model check reserves enough output for thinking before visible text. A reply that ends at its token budget is identified as incomplete; an unfinished tool call never runs.

Strict thinking APIs require private protocol state during a tool loop. Koder preserves that state in the live loop. Saved visible history deliberately omits private reasoning. When restoring such history, Koder projects a bounded quoted reference for the model rather than forging missing reasoning blocks. The saved conversation stays intact; quoted history cannot grant permission or replay an old tool action. A plain greeting remains isolated from old tasks.

Moonshot K3, Moonshot thinking and tool loops, DeepSeek thinking, Kimi Code models and membership.

Moonshot Responses setup, DeepSeek Responses setup.