SiliconFlow

💸$15.51 vs. $120.93. Same prompt. Same 3D game challenge. No manual code edits. GLM-5.3 completed the build at nearly 1/8 the cost of FABLE 5! Watch the side-by-side comparison—and decide for yourself. 👇 GLM-5.3 is now on SiliconFlow.

GLM-5.3 from @Zai_org is now open source — and live on SiliconFlow. 🚀 Day-0 support, as always. No extra scale. A lot more intelligence. 744B total / 40B active — the same base model as GLM-5.2, pushed much further through post-training. Competitive with Fable 5 and GPT-5.6 Sol across frontier coding and agentic benchmarks. SiliconFlow ranks #1 in GLM-5.2 token share on @OpenRouter. Now we’re bringing that experience to GLM-5.3. Build with GLM-5.3 now.

10% cheaper. 20× more TPM. 🔥 Your limits on Kimi K3 just got boosted on SiliconFlow. Starting now: 💸 10% OFF Kimi K3 ⚡ Default TPM: 100K → 2M New pricing: Cache input $0.30/M → $0.27/M Input $3.00/M → $2.70/M Output $15.00/M → $13.50/M More capacity. Lower cost. Available to everyone! Build with Kimi K3 on http://siliconflow.com 🚀

The hottest open-source models on OpenRouter. Now live on SiliconFlow. 🔥 Meet Hy4 preview: 770B total parameters ⚡ 49B active per token 🧠 1M context 🛠️ Built for coding, analysis, research & complex real-world work 🔓 Apache 2.0 With SiliconFlow, plug Hy4 preview straight into the agents and tools you already use: Claude Code. Codex. Cursor — and more. Your tools. Hy4 preview. Powered by SiliconFlow. Fast, reliable inference — ready for real-world workloads.

SiliconFlow Still Supports DeepSeek V4 Flash & V4 Pro — Keep Your Existing Workflows Running. DeepSeek’s V4 lineup is changing 👀 V4 Flash has been retired, with V4 Pro set to follow on Sep 14. Still using the original V4 models? No rush to say goodbye. V4 Pro 0813, V4 Pro (0424), and V4 Flash 0731 are all available on SiliconFlow. 🧵 With BYOK, your existing V4 workflows can keep running: Bring your SiliconFlow API key → pick the original V4 Flash / V4 Pro → keep building with the models already in your stack. Still need the original V4? You know where to find them :) https://www.siliconflow.com/models/deepseek-ai

@deepseek_ai just made Flash the new Pro. 🤯 DeepSeek-V4.1-Flash is live on SiliconFlow — Day 0. 552B MoE — ~8B active during prefill, ~16B during decode. ⚡ Faster inference, higher throughput 👁️ Native vision 🧠 1M context window 📉 ~4× smaller KV cache footprint vs. V4 Flash 📖 MIT licensed Run it on SiliconFlow — production-ready, high-throughput inference that keeps Flash fast at scale. Flash in name. Flagship in performance. Build with it → http://siliconflow.com

DeepSeek V4 Flash just got vision. 👀 DeepSeek-V4-Flash-Vision-Exp is now live on SiliconFlow — and available via @OpenRouter too. Built on V4 Flash, now with visual understanding for multimodal tasks and agents. Strong text capabilities. Stronger multimodal agent performance. 1M context. MIT licensed. Build with DeepSeek Vision on SiliconFlow. 🚀

Not every model call needs an answer. Sometimes, it just needs a decision. 👏 Welcome Kev-4B to SiliconFlow. Kev-4B is the open-source community edition of Jev, built on Qwen3.5-4B for structured decision-making. Route. Rank. Approve. Escalate — without another generated response. No deployment or adaptation needed. One SiliconFlow API key, and Kev is ready to plug into your workflow. Huge shoutout to @jaredpalmer for open-sourcing Kev. ❤️ Try Kev-4B on SiliconFlow. ⚡️

🚀 DeepSeek-V4.1-Flash-1017 just landed on SiliconFlow! Stronger coding. Faster agent runs. Lower cost per task. ⚡ What's new? • Agentic Coding: 77.0% on DeepSWE v1.1, with stronger end-to-end coding capabilities. • Vibe Coding: Turn ideas into working projects, from games to 3D apps. • Faster Execution: Get tasks done in less time, with fewer output tokens. • Better Performance per Dollar: Pushing the performance–cost frontier further across agent benchmarks. Give it a spin👇 https://cloud.siliconflow.com

🐰 Space Bunny just went open-weight! MiniMax-M3.1-Flash from @MiniMax_AI is now live on SiliconFlow — Day 0 of its open-weight release! 🚀 Stronger agents. Faster inference. Lower cost per task. ⚡ What's new? Agentic Coding: 73.7% on DeepSWE v1.1 Computer Use: 61.7 on OSWorld V2 Native Multimodal: Text, image & video inputs with 1M context Faster & More Efficient: Nearly 3× faster inference and 70% lower agentic task costs vs. M3 MSA Architecture: Sparse attention across all layers for greater efficiency More agent power. Less cost per task. Build with it → https://cloud.siliconflow.com/

Same prompt, different models. DeepSeek V4.1 Flash 1017 vs. GPT-6 Astra — both building the same interactive 3D world with Three.js. Nearly the same quality, if not better — at over 50% lower cost than Astra. Prompt in the comments. 💬 Here's the side-by-side comparison 👇