{"name":"m10dis","repo":"crgregersen/m10dis","url":"https://midorreal.com/project/crgregersen-m10dis","repository":"https://github.com/crgregersen/m10dis","description":"Qwen3.8-Flash-Next-FP8 (48-layer MoE, 512 experts) on one RTX 3070 Ti plus 24 Maxwell GPUs from six Tesla M10 cards. PCIe multicast, GPU page-table remapping, 96 tok/s decode.","facts":{"license":{"value":"MIT","source":"https://github.com/crgregersen/m10dis","checked":"20 Nov 2026"},"language":{"value":"Cuda","source":"https://github.com/crgregersen/m10dis","checked":"20 Nov 2026"},"releases":{"value":"0","source":"https://github.com/crgregersen/m10dis/releases","checked":"20 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/crgregersen/m10dis/graphs/contributors","checked":"20 Nov 2026"},"archived":{"value":"no","source":"https://github.com/crgregersen/m10dis","checked":"20 Nov 2026"}},"stars":41.0,"writeup":{"what_it_is":"m10dis is a research system that runs the Qwen3.8-Flash-Next-FP8 mixture-of-experts model across one RTX 3070 Ti and 24 Maxwell GPUs from six Tesla M10 cards. The RTX handles attention, linear attention, the router and hyper-connections under vLLM, while the M10 GPUs hold and compute the routed and shared experts, the n-gram embedding table and the LM head, connected through one PCIe switch.","audience":"Researchers and engineers with a similar single-machine PCIe multi-GPU setup","claims":[{"kind":"specific","claim":"Decode runs at 52.5 tokens per second for a single stream","excerpt":"| **52.5 tok/s** (single stream) |","status":"not_checked"},{"kind":"specific","claim":"Prefill runs at about 277 tokens per second using pipelined batches","excerpt":"**≈ 277 tok/s** (pipelined batches of 32 rows, two in flight)","status":"not_checked"},{"kind":"specific","claim":"Supports 94,631 tokens of KV cache with prefix caching off","excerpt":"94,631 tokens of KV cache (2.5 GiB, prefix caching off)","status":"not_checked"},{"kind":"specific","claim":"Each Maxwell GPU holds 32 of the 640 intermediate rows of every expert","excerpt":"holds **32 of the 640 intermediate rows of every expert**","status":"not_checked"},{"kind":"specific","claim":"Scores 90.98 on MMLU-Redux 5-shot on a subset of questions","excerpt":"**90.98 ± 0.56** on 2,628 of 5,330 questions","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-07"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":22,"reality":24,"gap":-2,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-04T12:21:58+00:00","text":"Repository created","source":"https://github.com/crgregersen/m10dis"}],"cite":"m10dis is a GitHub project at https://github.com/crgregersen/m10dis written mainly in Cuda, described there as \"Qwen3.8-Flash-Next-FP8 (48-layer MoE, 512 experts) on one RTX 3070 Ti plus 24 Maxwell GPUs from six Tesla M10 cards. PCIe multicast, GPU page-table remapping, 96 tok/s decode\".","attribution":"Data from Mid or Real, https://midorreal.com/project/crgregersen-m10dis, checked 22 November 2026. Attribution with a link is required.","still_real":null}