DeepSeek-V4.1-Flash-1017 is live on the Alibaba Cloud Token Plan. · 77.0 Pass@ 1 on DeepSWE v1.1, ~$0.20/task · 1M context, 384K output, multimodal · Same price as deepseek-v4.1-flash with off-peak discounts. Sign up and start building today!

Deepseek-v4.1-flash-1017 on 4x H200 power capped, no quantisation. I've been cranking on this thing for weeks now. Any enterprise with 120k USD is able to serve 16 workers concurrently at 150+ tok/s each

Deepseek v4.1 Flash 1017 has now generated more local tokens for me than any other model If there's not some new GLM very soon, it'll probably become my top used model of all time and then probably more tokens than all others combined. such a cool model

DeepSeek V4.1 Flash 1017 is now the #1# open model on Agent Arena leaderboard. Hosted 🏆 #1# open Agent Arena 💰 ~$0.19 median/task 🔓 MIT weights Localmaxxers are already running the same 1017 checkpoint on 2× RTX 3090s + 128GB DDR4. And getting ⚡ ~30 tps decode 🚀 301–470 tps prefill This is a monster model w/552B MoE backbone 💾 + ~196B Engram memory 📚 1M native context The local build squeezes the experts to ~2.03 bpw and uses exact routing, with the big Engram tables on storage. ⚠️ Those 2x3090s use an aggressive quant. It does not mean the local 2-bit version retains the exact Arena score. But that's still kind of nuts. Link to the 2x 3090's test in the ALT

🚀 DeepSeek V4.1 Flash 1017 now runs on just 2x DGX Sparks 🔥 TensorFold 2-node engine + LibertAI's EXL3 2.9 bpw pack: full 1M context, loads in 27s, read a 131K-token prompt at 2,059 tok/s Speed numbers dropping soon 👀 Repo 👇

🚀 DeepSeek V4.1 Flash 1017 now runs on TensorFold across 4 DGX Sparks ⚡️ 89.8 tok/s for 1 user (vLLM: 48.5) 🔥 202.9 tok/s with 6 users 🧠 LibertAI's 3.5bpw EXL3, every matrix 🛠️ No new kernels needed New default lane in the repo 👇

DeepSeek V4.1 Flash 1017 updated numbers 1 stream tok/s, old → new: 💻 Code 73.8 → 76.3 📊 Format 55.4 → 75.8 🧾 JSON 52.1 → 63.2 ➗ Math 50.9 → 56.2 🧠 Reasoning 37.8 → 55.6 ✍️ Prose 24.4 → 36.1 📝 Summary 25.7 → 35.8 📖 Narrative 24.9 → 31.7

DeepSeek V4.1 Flash 1017 on 4x DGX Spark (TP4) 🚀 Same 4 Sparks vs the old V4.1 Flash: ⚡ 6 streams: 131.9 → 165.7 tok/s throughput (+26%) 📥 93K token prompt: 78s → 42s 🧠 300K context, KV pool 1.07M → 1.52M tokens Per prompt numbers 👇