DeepSeek V4.1 Flash 1017 on four DGX Sparks: two quant recipes build voxel worlds
Same prompt in DeepSeek Harness compares Mia's FP8/MXFP4 checkpoint with bertholomus' EXL3 4.0bpw quant: 93.5 min and 258,931 tokens for an 8.5/10 world versus 49.7 min and 121,707 tokens for 7.5/10, at 82.8 vs 69.5 tok/s single-request.
Two DeepSeek-V4.1-Flash-1017 voxel worlds on four DGX Sparks, side by side: Mia's TensorFold recipe against bertholomus' recipe. Setup: – Same prompt, same harness (DeepSeek Harness 0.2.1-rc.2), thinking on, one try, no help – Mia: FP8/MXFP4 checkpoint as DeepSeek published it, TensorFold TP4, commit 6beaad9 with the 8 TF_DS_* switches her README table sets on – bertholomus: his EXL3 4.0 bpw quant, TensorFold TP4, commit 8f84e7a – Context 1,048,576 tokens on both, answer limit 262,144 tokens on both, the figure on the model card – My Sparks are cabled as a switchless ring Mia, world The Lantern Isles: – Agent time: 93.5 min – Tokens written: 258,931 – Browser check: 9.0/10 – My rating of the world: 8.5/10 bertholomus, world Aetherfall: – Agent time: 49.7 min – Tokens written: 121,707 – Browser check: 9.0/10 – My rating of the world: 7.5/10 Speed: – One request, code: 82.8 tok/s (bertholomus) against 69.5 tok/s (Mia) – Eight requests at once: 141.2 tok/s (bertholomus) against 139.0 tok/s (Mia) in total, nearly equal – Same speed test for both: model-card sampling, thinking off, 512 tokens – Mia's README uses other settings (greedy code, prose with thinking on): on her own method my ring gets 119.9 tok/s on code, hers is 117.4 Good to know: – I prefer the art style of bertholomus' world – The world from Mia's recipe has clearly more detail, size and complexity, and that is why I rated it higher – Narrowly rank 1 in the 4 SPARKS 512 GB class, mostly because of the world: Mia's recipe is rank 1 of 3, bertholomus' is rank 2 (as of Oct 26) – The browser check's orbit test was fixed and now measures camera movement. Both worlds were re-scored with the same fixed checker: Mia's gets 9/10, bertholomus' stays at 9/10. Both still lose the frame-rate point to the test's software rendering. Play both worlds in your browser, every number and the raw output: Mia's run: https://niklaslenz.com/v2/deepseek-v41-flash-4sparks-mia-tp4-1017/ bertholomus' run: https://niklaslenz.com/v2/deepseek-v41-flash-4sparks-bertholomus-tp4/ Mia's recipe and the exact setup: https://mia-ai.net/models/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold Recipes by @MiaAI_lab and @bertholomusai. TensorFold engine by @ashxhart. Model by @deepseek_ai.