how to use siliconflow with atlas

Export SILICONFLOW_CN_API_KEY, run `atlas models siliconflow-cn` to list the catalog, then pin `"model": "siliconflow-cn/Qwen/Qwen3-Coder-480B-A35B-Instruct"` in atlas.json. Atlas hits https://api.siliconflow.cn/v1 via @ai-sdk/openai-compatible.

is there a free model i can use as a small model in a coding agent

Yes. SiliconFlow serves Qwen3.5-4B free at $0.00/$0.00 per Mtok with a 262,144 token context. Set `"small_model": "siliconflow-cn/Qwen/Qwen3.5-4B"` in atlas.json to make title, summary, and subagent traffic free.

is siliconflow safe for enterprise code

For most enterprises, no. The SiliconFlow API terminates in mainland China at api.siliconflow.cn, which is a data-residency and latency decision, not just a pricing one. If China-hosted inference is not acceptable, use DeepInfra or Together AI for the same Qwen weights on US infrastructure.

what do the Pro/ prefixed models mean on siliconflow

The `Pro/` prefixed models are the production tier and cost more than the unprefixed community tier. Pro/Kimi-K2.5 is $0.45/$2.25 per Mtok, so the headline SiliconFlow price is not always the one you pay. Check the prefix in `atlas models siliconflow-cn`.

how much does siliconflow cost per million tokens

Qwen3-Coder-480B-A35B is $0.25/$1.00 per Mtok, Qwen3-Coder-30B-A3B is $0.07/$0.28, DeepSeek-V3.2 is $0.27/$0.42, Qwen3.5-397B-A17B is $0.29/$1.74, and Qwen3.5-4B is $0.00/$0.00.

what context window does siliconflow support

SiliconFlow reaches 1,049,000 tokens on GLM-5.2 and DeepSeek-V4-Pro, and 262,144 tokens across the Qwen3.5 line. The 262,144 token window is what the coding slot on Qwen3-Coder-480B-A35B realistically needs.

Atlas with SiliconFlow in 2026: Qwen3 Coder 480B at $0.25/$1.00 per Mtok

SiliconFlow is the cheapest place in the Atlas registry to run Chinese open weights, and the prices are not close: Qwen3-Coder-480B-A35B is $0.25/$1.00 per Mtok here against $2.00/$2.00 on Together AI. Inside Atlas, the terminal-native coding agent, SiliconFlow gives you a strong coding model and a genuinely free small_model at Qwen3.5-4B ($0.00/$0.00 per Mtok). The catch is in the hostname, api.siliconflow.cn, which is a data-residency decision before it is a pricing one.

Key takeaways

Qwen3-Coder-480B-A35B is $0.25/$1.00 per Mtok on SiliconFlow, 8x cheaper on input than the $2.00/$2.00 Together AI listing of identical FP8 weights.
Qwen3.5-4B is served free at $0.00/$0.00 per Mtok with a 262,144 token context, a genuinely usable small_model at zero marginal cost.
The SiliconFlow API terminates in mainland China at api.siliconflow.cn, which will disqualify it outright at most enterprises.
Context reaches 1,049,000 tokens on GLM-5.2 and DeepSeek-V4-Pro, and 262,144 tokens across the Qwen3.5 line.
The `Pro/` prefixed rows are the production tier and cost more: Pro/Kimi-K2.5 is $0.45/$2.25 per Mtok.
SiliconFlow carries models that barely appear elsewhere: Baidu ERNIE-4.5-300B-A47B at $0.28/$1.10, Tencent Hunyuan-A13B, inclusionAI Ling-flash-2.0, and ByteDance Seed-OSS-36B.

What is SiliconFlow best at inside Atlas?

SiliconFlow is best inside Atlas as the cheapest host for Chinese open weights in 2026. Qwen3-Coder-480B-A35B runs at $0.25/$1.00 per Mtok on SiliconFlow, which is 8x cheaper on input than the $2.00/$2.00 Together AI listing of the identical FP8 weights.

An 8x input spread on identical weights is not a rounding error, it is the difference between running Atlas on a 480B coding model casually and rationing it. SiliconFlow also carries Qwen3-Coder-30B-A3B at $0.07/$0.28 per Mtok and DeepSeek-V3.2 at $0.27/$0.42, plus models that barely appear elsewhere: Baidu ERNIE-4.5-300B-A47B at $0.28/$1.10, Tencent Hunyuan-A13B, inclusionAI Ling-flash-2.0, and ByteDance Seed-OSS-36B. Atlas lets you switch the active model and provider on the fly with favorites and recents, so trying ERNIE-4.5-300B-A47B against Qwen3-Coder-480B-A35B on the same task is one keystroke, not a new account.

Does SiliconFlow have a free model for Atlas?

Yes. SiliconFlow serves Qwen3.5-4B free at $0.00/$0.00 per Mtok with a 262,144 token context, which makes it a genuinely usable `small_model` in Atlas at zero marginal cost. Setting it means Atlas title, summary, and subagent traffic becomes free.

A free small_model matters more than it sounds. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and every subagent generates title and summary overhead. On a metered provider that overhead is a quiet tax on every session. With `"small_model": "siliconflow-cn/Qwen/Qwen3.5-4B"` in atlas.json, the tax goes to zero and only the build model on the coding slot actually bills. Qwen3.5-4B is not a coding model, and it does not need to be: it is doing titles and summaries, and it does that at a 262,144 token context for nothing.

What is the catch with SiliconFlow?

The SiliconFlow API terminates in mainland China, at api.siliconflow.cn. That is a data-residency and latency decision, not just a pricing one, and it will disqualify SiliconFlow outright at most enterprises in 2026 regardless of how good the $0.25/$1.00 per Mtok price on Qwen3-Coder-480B-A35B looks.

An Atlas coding agent sends repository context to whatever model you pin. With SiliconFlow, that context goes to api.siliconflow.cn. For an individual developer on a hobby project, that may be an acceptable tradeoff for an 8x input discount. For a company with data classification policies, it is usually a hard no before anyone looks at the price. The documented Atlas alternative is direct: if China-hosted inference is not acceptable, use DeepInfra or Together AI for the same Qwen weights on US infrastructure instead. You pay more and you keep the residency posture.

What are the Pro/ prefixed models on SiliconFlow?

The SiliconFlow `Pro/` prefixed models are the production tier, and they cost more than the unprefixed community tier. Pro/Kimi-K2.5 is $0.45/$2.25 per Mtok, for example, so the headline SiliconFlow price is not always the one you pay. Run `atlas models siliconflow-cn` and read the prefixes.

The two-tier catalog is the most common source of surprise on a SiliconFlow bill. The community rows are the ones that produce the eye-catching numbers, and the `Pro/` duplicates sitting next to them are the ones a production workload actually wants. Before you pin a model in atlas.json, decide which tier you are on and pin that exact row. Atlas is precise about model IDs, so `siliconflow-cn/Qwen/Qwen3-Coder-480B-A35B-Instruct` and a `Pro/` variant are different pins with different bills.

What context window does SiliconFlow give Atlas?

SiliconFlow reaches 1,049,000 tokens on GLM-5.2 and DeepSeek-V4-Pro, and 262,144 tokens across the Qwen3.5 line. A million-token window on GLM-5.2 is more than an Atlas session typically needs, because Atlas retrieves targeted code rather than pushing whole repositories into context.

Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and it indexes code by AST declarations using tree-sitter, not blind line windows. That means the practical value of a 1,049,000 token SiliconFlow window is headroom for a long build session, not a license to dump the repository into the prompt. The 262,144 token context across the Qwen3.5 line is more than sufficient for the coding slot, and Qwen3-Coder-480B-A35B at $0.25/$1.00 per Mtok is where most Atlas users on SiliconFlow will actually spend.

When should I pick a different provider than SiliconFlow?

Pick a different provider than SiliconFlow whenever China-hosted inference is not acceptable, since the API terminates at api.siliconflow.cn in mainland China. Atlas documents DeepInfra or Together AI as the alternative for the same Qwen3-Coder-480B-A35B weights on US infrastructure, at $2.00/$2.00 per Mtok on Together AI.

SiliconFlow is the clearest cost-versus-residency tradeoff in the 2026 Atlas registry. Qwen3-Coder-480B-A35B at $0.25/$1.00 per Mtok against $2.00/$2.00 on Together AI is a real 8x input saving on identical FP8 weights, and Qwen3.5-4B at $0.00/$0.00 is a free small_model nobody else matches. Those are strong reasons to choose SiliconFlow. The hostname is a strong reason not to. There is no clever configuration that resolves this: either your organization can send repository context to api.siliconflow.cn or it cannot, and the price only matters if the answer is yes.

Setup

01Export SILICONFLOW_CN_API_KEY. Atlas hits https://api.siliconflow.cn/v1 via @ai-sdk/openai-compatible.
02Run `atlas models siliconflow-cn` to list the catalog, and note the `Pro/` rows are the production-tier duplicates that cost more.
03Pin `"model": "siliconflow-cn/Qwen/Qwen3-Coder-480B-A35B-Instruct"` in atlas.json for the coding slot at $0.25/$1.00 per Mtok.
04Set `"small_model": "siliconflow-cn/Qwen/Qwen3.5-4B"` to make Atlas's title, summary, and subagent traffic free at $0.00/$0.00 per Mtok.
05If China-hosted inference is not acceptable, use DeepInfra or Together AI for the same Qwen weights on US infrastructure instead.

Frequently asked questions

how to use siliconflow with atlas: Export SILICONFLOW_CN_API_KEY, run `atlas models siliconflow-cn` to list the catalog, then pin `"model": "siliconflow-cn/Qwen/Qwen3-Coder-480B-A35B-Instruct"` in atlas.json. Atlas hits https://api.siliconflow.cn/v1 via @ai-sdk/openai-compatible.
what is the cheapest way to run qwen3 coder 480b: SiliconFlow, at $0.25/$1.00 per Mtok for Qwen3-Coder-480B-A35B. That is 8x cheaper on input than the $2.00/$2.00 Together AI listing of the identical FP8 weights. The tradeoff is that the API terminates at api.siliconflow.cn.
is there a free model i can use as a small model in a coding agent: Yes. SiliconFlow serves Qwen3.5-4B free at $0.00/$0.00 per Mtok with a 262,144 token context. Set `"small_model": "siliconflow-cn/Qwen/Qwen3.5-4B"` in atlas.json to make title, summary, and subagent traffic free.
is siliconflow safe for enterprise code: For most enterprises, no. The SiliconFlow API terminates in mainland China at api.siliconflow.cn, which is a data-residency and latency decision, not just a pricing one. If China-hosted inference is not acceptable, use DeepInfra or Together AI for the same Qwen weights on US infrastructure.
what do the Pro/ prefixed models mean on siliconflow: The `Pro/` prefixed models are the production tier and cost more than the unprefixed community tier. Pro/Kimi-K2.5 is $0.45/$2.25 per Mtok, so the headline SiliconFlow price is not always the one you pay. Check the prefix in `atlas models siliconflow-cn`.
how much does siliconflow cost per million tokens: Qwen3-Coder-480B-A35B is $0.25/$1.00 per Mtok, Qwen3-Coder-30B-A3B is $0.07/$0.28, DeepSeek-V3.2 is $0.27/$0.42, Qwen3.5-397B-A17B is $0.29/$1.74, and Qwen3.5-4B is $0.00/$0.00.
what context window does siliconflow support: SiliconFlow reaches 1,049,000 tokens on GLM-5.2 and DeepSeek-V4-Pro, and 262,144 tokens across the Qwen3.5 line. The 262,144 token window is what the coding slot on Qwen3-Coder-480B-A35B realistically needs.

Try Atlas in your terminal

The terminal-native AI coding agent. Free core, single binary.

Install Atlas

Related guides

Rename a Symbol Across the Repo with Atlas in 2026

How to rename a symbol across a repo with Atlas in 2026: findReferences gets the true reference set, grep catches strings and docs, and edit refuses ambiguous matches.

Atlas vs Mistral Vibe for Code: Terminal AI Coding Agents in 2026

Compare Atlas and Mistral Vibe for Code in 2026. Atlas offers terminal-native TUI, explicit diffs, and BYO models. Mistral Vibe for Code provides a four-model stack, multi-platform access, and EU data sovereignty.

Atlas for Rust in 2026

Adopt Atlas, the terminal-native AI coding agent, for Rust development in 2026. Tackle borrow checker errors and clippy lints with Atlas's secure, approval-gated assistance.

Atlas for Elixir in 2026

Adopt Atlas, the terminal-native AI coding agent, for Elixir development in 2026. Enhance productivity with deep code understanding, safety features, and direct integration into mix projects and OTP applications.

Atlas for FastAPI in 2026

Atlas is a terminal-native AI coding agent for FastAPI in 2026. It reads Pydantic v2 models and Depends(), then runs pytest behind a permission prompt.

Atlas for Fortran: fpm.toml, Explicit Interfaces, and fprettify in 2026

Atlas is a terminal-native AI coding agent for Fortran in 2026. It reads modules, explicit interfaces, and intent declarations, runs fpm test behind a prompt, and runs fprettify.

Atlas vs Pieces for Developers: AI Tools for Developers in 2026

Comparing Atlas, a terminal-native AI coding agent, with Pieces for Developers, an OS-level memory layer, for developers in 2026. Evaluate code generation, safety, and context management.

Atlas vs Blackbox AI: Choosing Your AI Coding Agent in 2026

Comparing Atlas, the terminal-native AI coding agent, with Blackbox AI, a VS Code agent with 4.7 million installs, for developers in 2026. Evaluate features, pricing, and workflow.

Browse this resource hub