Skip to main content

SiliconFlow Provider Guide

SiliconFlow is a Tier-2 catalog provider: its integration is one JSON file (src/lib/providers/catalog/siliconflow.json) rather than hand-written code. That file is the source of truth for the values on this page.

Verification status: this entry is docs-verified only, not yet live-verified. SiliconFlow's GET https://api.siliconflow.com/v1/models answers HTTP 401 without a key, so the model ids come from SiliconFlow's public docs and model pages and the roster has not been checked. No account was created and no API key was used to build the entry. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: siliconflow
  • Protocol: POST /chat/completions with bearer auth — the chat-completions API reference lists the server https://api.siliconflow.com/v1 and POST /chat/completions with bearer auth. The Text Generation page says "You can make end-to-end API requests using the OpenAI SDK" and the Quick Start says "The platform currently supports most OpenAI-related parameters."
  • Base URL: https://api.siliconflow.com/v1
  • Default model: deepseek-ai/DeepSeek-V4-Pro
  • Models in catalog: 12 — ids taken from the model enums of the chat-completions API reference and matching the ids shown on the model pages (linked under See also); roster not verified
  • Messages array: the API reference's request schema for messages shows minItems: 1 and maxItems: 10
  • Streaming: supported — the API reference documents stream: "If set, tokens are returned as Server-Sent Events as they are made available." The Stream Mode page shows stream=True with the OpenAI library
  • Tool calling: model-dependent — the Function Calling page says "The Function Calling feature allows the model to call external tools to enhance its capabilities." and the API reference documents tools ("A max of 128 functions are supported."). The model pages show a Tools field as "Supported" or "Not supported" per model: "Supported" on the 12 catalog models and "Not supported" on openai/gpt-oss-120b. None of it was exercised
  • Tools while streaming: not declared (false)
  • Structured output: not declared (false) — Structured Outputs reads "Not supported" on the 12 catalog model pages. The JSON Mode page documents response_format={"type": "json_object"} and says "Currently, online models, except for the DeepSeek R1 series and V3 models, support the above parameter." JSON Mode reads "Supported" on 10 of the 12 catalog model pages, among them deepseek-ai/DeepSeek-V3.2, deepseek-ai/DeepSeek-V4-Pro and deepseek-ai/DeepSeek-V4-Flash, and "Not supported" on zai-org/GLM-5.1 and zai-org/GLM-5. The entry sets no responseFormatDowngrade quirk
  • Structured output + tools together: not declared (false) — no combined probe was run without credentials
  • Embeddings: not declared (false) — SiliconFlow documents a POST /embeddings route (Create embeddings). src/lib/providers/configuredOpenAICompat.ts, which the catalog entries use, contains no embed()/embedMany() implementation, so this flag tracks that, not the vendor's API surface
  • Thinking: not declared (false) — the API reference documents enable_thinking ("Switches between thinking and non-thinking modes. Default is True.") with the models it supports, thinking_budget, and a reasoning_content response field. This entry sends none of them
  • Vision: true for six catalog models whose model page shows Support image input "Supported": moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.5, Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B, Qwen/Qwen3-VL-32B-Instruct and google/gemma-4-31B-it. models.visionModel is Qwen/Qwen3-VL-32B-Instruct, the default of the API reference's VLM request schema, which lists it in its model enum. For moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.5, Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B and google/gemma-4-31B-it the flag rests on the model page value; image input was not exercised
  • Billing: free-tier — see Billing below
  • Key format: not validated by NeuroLink (apiKeyFormat: null)

Quick Start​

1. Get an API key​

  1. Visit: https://cloud.siliconflow.com/ and create a SiliconFlow account (https://docs.siliconflow.com/en/userguide/quickstart)
  2. Open the API Keys page at https://cloud.siliconflow.com/account/ak, click Create API Key and copy the key (https://docs.siliconflow.com/en/userguide/quickstart)
  3. Billing as the vendor words it on https://www.siliconflow.com/pricing: "Flexible token pricing, high usage limits, and postpaid billing—plus $1 in free credits to get you started!"
  4. Set SILICONFLOW_API_KEY in your .env file

2. Configure​

export SILICONFLOW_API_KEY=your-api-key
export SILICONFLOW_MODEL=deepseek-ai/DeepSeek-V4-Pro   # optional — overrides the default model
export SILICONFLOW_BASE_URL=https://api.siliconflow.com/v1   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "siliconflow",
  model: "deepseek-ai/DeepSeek-V4-Pro",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider siliconflow

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "siliconflow",
  credentials: { siliconflow: { apiKey: process.env.SILICONFLOW_API_KEY } },
});

Billing​

SiliconFlow's pricing page says: "Flexible token pricing, high usage limits, and postpaid billing—plus $1 in free credits to get you started!" Its FAQ answers "Are there any minimum commitments?" with "No, there are no minimum commitments. You only pay for what you use, and you can start with $1 in free credits." The entry records free-tier.

The Error Handling page says of HTTP 403: "The most common reason is that the model requires real-name authentication."

Models​

ModelContextMax tokensVisionToolsJSON Mode$/M in · out (cache read)Notes
deepseek-ai/DeepSeek-V4-Pro ⭐1049K393KnoSupportedSupported$1.50162 / $3.135 ($0.135)NeuroLink default. About text begins "DeepSeek-V4-Pro is DeepSeek's flagship open-source MoE model".
deepseek-ai/DeepSeek-V4-Flash1049K393KnoSupportedSupported$0.13 / $0.28 ($0.028)About text begins "DeepSeek-V4-Flash is DeepSeek's latest open-source MoE model".
deepseek-ai/DeepSeek-V3.2164K164KnoSupportedSupported$0.27 / $0.42 ($0.135)About text begins "DeepSeek-V3.2 is a model that harmonizes high computational efficiency with superior reasoning and agent performance".
zai-org/GLM-5.1205K131KnoSupportedNot supported$1.19 / $3.74 ($0.6)About text begins "GLM-5.1 is Z.ai's next-generation flagship model built for agentic engineering".
zai-org/GLM-5205K131KnoSupportedNot supported$0.95 / $2.55 ($0.2)About text begins "GLM-5 is a next-generation open-source model for complex systems engineering and long-horizon agentic tasks".
moonshotai/Kimi-K2.6262K262KyesSupportedSupported$0.77 / $3.4 ($0.14)About text begins "Kimi K2.6 is an open-source, native multimodal agentic model by Moonshot AI".
moonshotai/Kimi-K2.5262K262KyesSupportedSupported$0.45 / $2.25 ($0.07)About text begins "Kimi K2.5 is an open-source, native multimodal agentic model".
Qwen/Qwen3.6-27B262K262KyesSupportedSupported$0.3 / $3.2About text begins "Qwen3.6-27B is the first open-weight small-to-mid-sized dense model in the Qwen3.6 series".
Qwen/Qwen3.6-35B-A3B262K262KyesSupportedSupported$0.24 / $1.8 ($0.15)About text begins "Qwen3.6-35B-A3B is a large language model from Alibaba's Qwen3.6 series".
Qwen/Qwen3-32B131Knot setnoSupportedSupported$0.14 / $0.57About text begins "Qwen3-32B is the latest large language model in the Qwen series with 32.8B parameters".
Qwen/Qwen3-VL-32B-Instruct262K262KyesSupportedSupported$0.2 / $0.6About text begins "Qwen3-VL is the vision-language model in the Qwen3 series".
google/gemma-4-31B-it262K262KyesSupportedSupported$0.75 / $1.0 ($0.25)About text begins "Gemma 4 31B is Google DeepMind's latest open-source model".

Context, max tokens, prices and the Tools, JSON Mode and Support image input values come from the model pages at https://www.siliconflow.com/models/<slug> (retrieved 2026-09-29). The Context and Max tokens columns show the page's "K" values; the catalog records them as thousands (1049K is stored as 1,049,000). Prices are the page's Input Price, Output Price and Cache Read figures, shown with a $ sign per "/ M Tokens". The pricing page shows the same context length and prices for deepseek-ai/DeepSeek-V4-Pro, zai-org/GLM-5.1, zai-org/GLM-5, moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.5, Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B and google/gemma-4-31B-it, under the names without the organisation prefix. The 12 ids are copied from the model enums of the API reference and match the ids on their model pages.

Qwen/Qwen3-32B has no maxOutputTokens in the catalog. Its model page shows Max Tokens "131K", and the Reasoning page lists the Qwen3 Series with Maximum Response Length 8192, Maximum Reasoning Chain Length 32768 and Maximum Context Length 131072.

models.defaultContextWindow (32,768) and models.defaultMaxOutputTokens (4,096) are placeholders for ids outside the catalog that the vendor does not publish. The status of the catalog models is production, a value the schema requires, not a SiliconFlow statement.

Fallback order when the default is unavailable: deepseek-ai/DeepSeek-V4-Flash → deepseek-ai/DeepSeek-V3.2 → zai-org/GLM-5.1 → moonshotai/Kimi-K2.6 → Qwen/Qwen3-32B. The order is NeuroLink's choice, not a vendor ranking. The runtime fallback model name the loader derives (fallbacks[1]) is deepseek-ai/DeepSeek-V3.2.

Reasoning fields​

The API reference documents enable_thinking ("Switches between thinking and non-thinking modes. Default is True.") for a list of models that includes deepseek-ai/DeepSeek-V3.2, thinking_budget ("Maximum number of tokens for chain-of-thought output.") and a reasoning_content field on the response message. The Reasoning page says: "reasoning_content: Reasoning chain content, at the same level as content." This catalog entry does not send enable_thinking or thinking_budget.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for SiliconFlow — docs-verified only, not live-verified:

ProbeResult
Rosterunauthenticated GET https://api.siliconflow.com/v1/models answers HTTP 401 without a key (2026-09-29), so model ids are taken from the API reference and the model pages (retrieved 2026-09-29); roster not verified
Auth-failure shapeunauthenticated GET https://api.siliconflow.com/v1/models returned HTTP 401 with body {"code":30014,"data":null,"message":"Token is invalid."} and header www-authenticate: Bearer realm="siliconflow", error="invalid_token", error_description="Token is invalid." (2026-09-29). An unauthenticated GET https://api.siliconflow.com/v1/chat/completions returned HTTP 404 with content-type: text/plain and the body Not Found. No POST was sent, so the 401 rule rests on the GET above, and the 403, 429 and model-not-found rules on the Error Handling page
Billingpublic pricing page statements quoted in Billing above, 2026-09-29
Tools / structured outputfunction calling is documented on the Function Calling page and response_format on the JSON Mode page; nothing was exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=siliconflow with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

Causes and fixes in quotation marks are SiliconFlow's wording, with the page named.

SymptomCauseFix
HTTP 401"API Key is not properly set." (Error Handling)"Verify the API Key" (Text Generation, error code table)
HTTP 403"The most common reason is that the model requires real-name authentication." (Error Handling)"For other cases, refer to the error message (message)." (Error Handling)
HTTP 429"Rate limits triggered." (Error Handling)"Refer to the error message (message) to determine whether the issue is related to RPM / RPD / TPM / TPD / IPM / IPD." (Error Handling); "Implement exponential backoff retry mechanism" (Text Generation, error code table)
HTTP 503 or 504"Generally caused by high system load." (Error Handling)"You can try again later." and "you can try using streaming output" (Error Handling)
HTTP 400 with code 20012The Error Handling page lists 400 as "Incorrect parameters." and shows the message "Model does not exist. Please check it carefully.""Refer to the error message (message) to correct invalid request parameters." (Error Handling)
max_tokens set to the context length"As some services are still being updated, avoid setting max_tokens to the window’s upper bound; reserve ~10k tokens as buffer for input and system overhead." (API reference)"Ensure that input tokens + max_tokens do not exceed the model’s context window." (same page)

See also​