DeepSeek V4.1 Flash builds a voxel sky world on 4x DGX Spark in 93.5 min, rated 8.5/10
On four DGX Sparks in TP4, V4.1 Flash 1017 generated an 875K-triangle floating-isle world in one Harness run, scoring 9/10 in the browser check at about 70 tok/s decode.
🚨 A new #1 in the 512GB category 🚨 DeepSeek-V4.1-Flash-1017 on four DGX Sparks built The Lantern Isles, a voxel sky world, in 93.5 min, with @MiaAI_lab's TensorFold recipe. I rated it 8.5 out of 10. By far the best world I have seen a model generate locally, and the first one that feels close to frontier quality. The amount of textures and the care for the small details separate itself from the others. Setup: – Recipe: Mia's, commit 6beaad9, with the 8 TF_DS_* switches her README table sets on; one model split across all four Sparks (TP4), 16 streams – Context: 1,048,576 tokens – Agent: DeepSeek Harness 0.2.1-rc.2, one prompt, one try, no help, thinking on – Sampling from the DeepSeek model card: temperature 1.0, top_p 1.0, top_k 20 Speed: – Decode, code: 69.5 tok/s (separate speed run, thinking off) – Decode, prose: 70.2 tok/s – 500 words: about 10.0 s – Eight requests at once: 139.0 tok/s in total The world: – Built in 93.5 min, 258,931 tokens – Browser check: 9.0/10 – Floating isles with a windmill, a village of lit houses, waterfalls, a hot-air balloon, a great lantern tree and its own camera tour with named chapters – 875,164 triangles in the scene Good to know: – In the 4 SPARKS 512 GB class, this run is rank 1 of 3 as of Oct 26. Narrowly rank 1, mostly because of the world. – After 90 minutes the harness told the model to finish, and it was done at 93.5. The browser check took 1 point off for the frame-rate test (under 1 fps, it renders in software). – The browser check's orbit test was fixed and now measures camera movement. Re-scored with the fixed checker, this world gets 9/10. – The recipe's default answer budget is 32,768 tokens, the model card allows 262,144, so I raised the limit to that. – On her own README method my ring gets 119.9 tok/s on code, hers is 117.4. Her README lists 165.3 tok/s at eight requests (prose, thinking on, temperature 0.7); my speed test runs thinking off at the model card's sampling. One thing I want to mention separately: I've tested around 40 recipes in the last week. And although Mia consistently holds the top spot, the one thing I appreciate the most is how well and easy the setup is. It happens quite often that I have to iron out a few bugs first, when I am testing a new recipe. Not here. Play The Lantern Isles in your browser, every number and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-4sparks-mia-tp4-1017/ Mia's recipe and the exact setup: https://mia-ai.net/models/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold Recipe by @MiaAI_lab. TensorFold engine by @ashxhart. Model by @deepseek_ai.