# Atlas with SiliconFlow in 2026 > SiliconFlow offers Qwen3-Coder-480B-A35B at an impressive $0.25 per million input tokens, making it the cheapest option for these weights. In 2026, Atlas users seeking the most cost-effective access to Chinese open-weights models will find SiliconFlow an unparalleled choice. It offers Qwen3-Coder-480B-A35B for just $0.25 per million input tokens, an 8x saving compared to other providers. SiliconFlow also provides a free Qwen3.5-4B model with a 262,144 token context window, ideal for Atlas's small_model tasks. ## Key takeaways - SiliconFlow offers Qwen3 Coder 480B A35B at $0.25/$1.00 per Mtok, which is 8x cheaper on input than Together AI. - Qwen3.5-4B is served free at $0/$0 per Mtok with a 262,144 token context, making it ideal for Atlas's `small_model`. - The API endpoint is api.siliconflow.cn in mainland China, posing data residency and potential latency concerns. - SiliconFlow carries models that barely appear elsewhere, including Baidu ERNIE-4.5-300B-A47B and Tencent Hunyuan-A13B. - Some models, like GLM-5.2 and DeepSeek-V4-Pro, offer a massive 1,049,000 token context window. - Pro/ prefixed models are the production tier and cost more than the unprefixed community tier. ## Why choose SiliconFlow for Atlas in 2026? For Atlas users in 2026, SiliconFlow stands out as the most economical provider for Chinese open-weights models, offering Qwen3 Coder 480B A35B at just $0.25 per million input tokens. This represents an 8x cost reduction compared to other services for identical FP8 weights. SiliconFlow is the cheapest place in the registry to run Chinese open weights, providing significant cost savings for Atlas users. For instance, Qwen3 Coder 480B A35B is available for $0.25 per million input tokens and $1.00 per million output tokens, which is 8x cheaper on input than the $2.00/$2.00 Together AI listing for the identical FP8 weights. Beyond cost, SiliconFlow carries models that barely appear elsewhere, such as Baidu ERNIE-4.5-300B-A47B, Tencent Hunyuan-A13B, inclusionAI Ling-flash-2.0, and ByteDance Seed-OSS-36B. Furthermore, SiliconFlow serves Qwen3.5-4B for free at $0/$0 per million tokens, offering a 262,144 token context window. This free model is a genuinely usable `small_model` for Atlas, allowing it to handle tasks like generating titles, summaries, and managing subagent traffic without incurring marginal costs. ## What are the pricing and context window tradeoffs with SiliconFlow? SiliconFlow offers a range of models with varying costs and context windows, including a massive 1,049,000 tokens for GLM-5.2 and DeepSeek-V4-Pro, and 262,144 tokens for the Qwen3.5 line. The Qwen3-Coder-480B-A35B model is priced at $0.25/$1.00 per Mtok, but users must be aware of the Pro/ tier models, which cost more. SiliconFlow provides a diverse catalog of models with distinct pricing and context window capabilities. For high-context tasks, models like GLM-5.2 and DeepSeek-V4-Pro offer an expansive 1,049,000 token context window. The Qwen3.5 line, including Qwen3.5-397B-A17B and Qwen3.5-4B, provides a substantial 262,144 token context window. Pricing for key models includes Qwen3-Coder-480B-A35B at $0.25/$1.00 per Mtok, Qwen3-Coder-30B-A3B at $0.07/$0.28 per Mtok, DeepSeek-V3.2 at $0.27/$0.42 per Mtok, and Qwen3.5-397B-A17B at $0.29/$1.74 per Mtok. A significant advantage is the Qwen3.5-4B model, which is served free at $0.00/$0.00 per Mtok. However, a tradeoff to note is that `Pro/` prefixed models, which are the production tier, cost more than the unprefixed community tier. For example, Pro/Kimi-K2.5 is priced at $0.45/$2.25 per Mtok, meaning the headline community tier price is not always the one you will pay for production use cases. ## What are the data residency and latency considerations for SiliconFlow? A critical consideration for Atlas users in 2026 is SiliconFlow's API endpoint, which terminates in mainland China at api.siliconflow.cn. This geographical location presents significant data-residency and potential latency implications, which may disqualify its use for many enterprise environments. The primary tradeoff when considering SiliconFlow for Atlas is its API endpoint, api.siliconflow.cn, which is located in mainland China. This is not merely a pricing decision but a crucial data-residency and latency consideration. For organizations with strict compliance requirements or those operating outside of Asia, the data residency in mainland China will likely disqualify SiliconFlow outright. Depending on your geographical location, connecting to an API in China may also introduce noticeable latency, impacting the responsiveness of Atlas. Developers must weigh the substantial cost savings against these operational and compliance factors. If China-hosted inference is not acceptable for your use case, Atlas users can opt for alternative providers like DeepInfra or Together AI, which offer the same Qwen weights on US infrastructure, albeit at a higher cost. ## When should I choose a different model or provider over SiliconFlow? While SiliconFlow offers compelling cost advantages, particularly with Qwen3 Coder 480B A35B at $0.25 per Mtok, it is not suitable for every Atlas workflow. If your organization has strict data residency requirements that prohibit data processing in mainland China, or if low latency from non-Chinese regions is paramount, you should consider alternatives. You should choose a different model or provider over SiliconFlow if your enterprise or project has strict data residency policies that forbid data processing in mainland China. The api.siliconflow.cn endpoint means all data processed by SiliconFlow models will reside in China, which is a non-starter for many organizations. Similarly, if low latency from regions outside of China is a critical performance requirement for your Atlas workflows, the geographical distance to SiliconFlow's servers might introduce unacceptable delays. While SiliconFlow offers unique access to models like Baidu ERNIE-4.5-300B-A47B and the cheapest Qwen weights, if these specific models are not a priority, or if US-based infrastructure is a must, providers like DeepInfra or Together AI offer the same Qwen weights on US infrastructure. Additionally, if you require the production-tier `Pro/` models, their increased cost might diminish the overall savings, making other providers more competitive depending on your specific usage patterns. ## Setup 1. Export your SiliconFlow API key: `export SILICONFLOW_CN_API_KEY="your_api_key_here"` 2. List the available models from SiliconFlow: `atlas models siliconflow-cn` (note the `Pro/` rows for production tiers) 3. Pin your preferred coding model in `atlas.json`, for example: `"model": "siliconflow-cn/Qwen/Qwen3-Coder-480B-A35B-Instruct"` 4. Set a free small model for background tasks in `atlas.json`: `"small_model": "siliconflow-cn/Qwen/Qwen3.5-4B"` 5. If China-hosted inference is not acceptable, use DeepInfra or Together AI for the same Qwen weights on US infrastructure instead. ## FAQ ### What is the primary advantage of using SiliconFlow with Atlas? SiliconFlow provides the cheapest access to Chinese open-weights models in 2026, with Qwen3 Coder 480B A35B costing $0.25 per million input tokens, an 8x saving over other providers for identical weights. ### Can I use SiliconFlow for free with Atlas? Yes, SiliconFlow offers Qwen3.5-4B at $0.00/$0.00 per Mtok, providing a genuinely usable `small_model` with a 262,144 token context window for Atlas's title, summary, and subagent traffic without marginal cost. ### What are the context window sizes available on SiliconFlow? SiliconFlow offers models with context windows up to 1,049,000 tokens for models like GLM-5.2 and DeepSeek-V4-Pro, and 262,144 tokens for models on the Qwen3.5 line. ### What is the main tradeoff when using SiliconFlow? The primary tradeoff is that SiliconFlow's API terminates in mainland China (api.siliconflow.cn), which introduces data-residency and latency considerations that may disqualify it for many enterprises. ### Are all models on SiliconFlow priced the same? No, pricing varies significantly. For example, Qwen3-Coder-480B-A35B is $0.25/$1.00 per Mtok, while `Pro/` prefixed models are production tier and cost more, such as Pro/Kimi-K2.5 at $0.45/$2.25 per Mtok. ### How do I configure Atlas to use SiliconFlow? You need to export your `SILICONFLOW_CN_API_KEY`, then run `atlas models siliconflow-cn` to see the catalog. Pin your desired model, like `"siliconflow-cn/Qwen/Qwen3-Coder-480B-A35B-Instruct"`, in your `atlas.json` for the coding slot. ### What if my organization cannot use China-hosted inference? If China-hosted inference is not acceptable due to data residency or latency, you should use providers like DeepInfra or Together AI, which offer the same Qwen weights on US infrastructure instead. --- Canonical HTML: https://seashell.sh/resources/models/siliconflow Source of truth: aeo_pages row `/resources/models/siliconflow` (segment: Models) (this file is generated from it, never hand-edited). Licence: SeaShell is proprietary with a free core. It is not open source and there is no public source repository.