DEPLOYDatabase

Company

SiliconFlow

SiliconFlow (硅基流动) is a Beijing-based AI infrastructure startup founded in August 2023 by Yuan Jinhui, the same engineer who previously built the OneFlow…

Founded
2023
Status
Private

The verified answer

As of Oct 2026, DEPLOY verifies SiliconFlow: $280.0M raised. 6 sources back the record.

Overview

SiliconFlow (硅基流动) is a Beijing-based AI infrastructure startup founded in August 2023 by Yuan Jinhui, the same engineer who previously built the OneFlow deep-learning framework. Its flagship product, SiliconCloud, is a serverless inference API that hosts more than 200 open and open-weights models, including DeepSeek, Qwen, GLM, Kimi, Meta Llama derivatives, FLUX image models, and Wan2.2 video generation. The platform bills per token rather than per GPU-hour and positions itself as an OpenAI-compatible substitute in the Chinese market.

Unlike Western neoclouds such as CoreWeave or Lambda, SiliconFlow does not publicly disclose owning datacenter capacity or GPU fleets of its own. It operates as an inference-optimization layer that runs across partner infrastructure, and its public materials emphasize a proprietary inference engine plus open-source projects (OneDiff for diffusion, BizyAir for scalable inference) rather than a physical footprint. Chinese-market disclosures rank it fourth in national token throughput for 2025 at roughly 1.5 percent market share, and first among independent (non-hyperscaler) token suppliers.

The company scaled quickly on the back of the DeepSeek wave. After an Innovation Works-led angel in 2024 and a Pre-A led by China Innovation Funding, it raised a Series A of hundreds of millions of RMB in June 2025 with Meituan taking a strategic stake, then closed a Series B in June 2026 of more than 2 billion RMB at a post-money valuation of about 7.74 billion RMB (roughly USD 1.08 billion), and filed for a Hong Kong Stock Exchange listing the same month.

Reality check

LLM pricing (company-claimed)

Per 1M input tokens: DeepSeek $0.13 to $1.50, Qwen $0.10 to $0.30, Kimi $0.45 to $3.00.

Multimodal pricing (company-claimed)

Images $0.014 to $0.08 each on FLUX.1; video $0.29 per clip on Wan2.2.

Verified record

Verified deployments
None on file
Active incidents
None on file

Key facts

Founder background

Founded by Yuan Jinhui, previously founder of the OneFlow deep learning framework.

Core product

SiliconCloud, a serverless inference API hosting 200+ LLMs and multimodal models behind an OpenAI-compatible interface.

Positioning

Inference layer on partner infrastructure; owned GPU counts, vendor mix and data center megawatts are not disclosed.

Market position (reported, 2025)

Ranked #4 in China by token throughput at about 1.5% share, and #1 among independent suppliers.

Models served

DeepSeek (including V4-Flash), Qwen, GLM-5.2, Kimi, Gemma, FLUX.1 and Wan2.2, plus TTS and ASR audio models.

Open-source projects

OneDiff (diffusion inference engine) and BizyAir (scalable inference runtime).

Series B size (reported, Jun 2026)

Over RMB 2B (about $280M) at a RMB 7.74B (about $1.08B) post-money valuation; no single lead named.

Series B backers (reported)

Trip.com, Jinko, Kingdee, China Unicom, Biren, NIO, SenseTime, Giant Interactive and state-backed funds, per 36Kr.

Earlier funding (reported)

Angel round (2024) led by Innovation Works; Pre-A (2024) led by China Innovation Funding; Series A (Jun 2025) of hundreds of millions of RMB.

Other backers (reported)

Meituan took a strategic stake in 2025; Huachuang Capital, Puhua Capital and Yaotu Capital also invested.

IPO filing (Jun 2026)

Filed for a Hong Kong Main Board listing under Chapter 18C on Jun 30 2026, with Huatai and Haitong as joint sponsors.

Data & sources

Web sources

6

6 sources backing this record.View all →

Current leadership (1)

Founders (1)

See all founders across the registry →

Safety record

No incidents on record for SiliconFlow.

Only active incidents are counted. Retracted incidents are excluded from this summary but remain reachable at their canonical URLs.

Full safety record: incidents, sourcing, and exposure data →