Sparks Update: MiniMax-M3.1-Flash on Two Sparks, and DeepSeek-1017 From Two to Four
The first two-Spark MiniMax numbers arrived overnight, the pack billed for two Sparks did not fit, and DeepSeek-1017 builders posted an engine comparison and a hybrid rig.
Yesterday's Sparks Update ended with two-Spark MiniMax builds waiting on their first numbers. Overnight they arrived, along with new DeepSeek-1017 builds on two, three and four Sparks. Every number below names its build and links its source.
At a glance
- MiniMax-M3.1-Flash runs on two Sparks. A 2.15 bpw EXL3 pack at TP2 measures 30.4 tok/s on code and 19.8 on prose at one stream, with 300K context.
- The pack billed for two Sparks did not fit. The 4.26 bpw build left 1.72 GiB free per node before any cache.
- Three Sparks, two engines. On DeepSeek-1017, TensorFold ran 108.6 tok/s at four sessions against 68.3 for SGLang, one builder reports.
- Two Sparks plus two RTX 3090s. Moving DeepSeek's drafter and Engram tables to an x86 server gives 49.7 to 66.2 tok/s per request.
- Ling 3.1 Flash on four Sparks took 12 hours 26 minutes on one prompt. The same prompt took 57 minutes on an API.
- Still missing: speeds for DeepSeek's official weights on four Sparks, and any first-hand report from an eight-Spark or 64 GB Spark owner.
MiniMax-M3.1-Flash on two Sparks
@ViC305 released two SAGE EXL3 quants of MiniMax's 428B model on Friday afternoon. He measured each against the original checkpoint with teacher-forced evaluation:
| Pack | Size | Target | Top-1 token agreement | Mean KL divergence |
|---|---|---|---|---|
| 2.15 bpw | 108.64 GiB | One Spark | 96.89% | 0.02528 |
| 4.26 bpw | 209.88 GiB | Two Sparks | 98.44% | 0.00592 |
He calls these quantization-fidelity measurements, not task pass rates. His one-Spark vLLM recipe works but reaches only 32K context with a quantized KV cache. He asked two-Spark owners to build the TP2 recipe.
@Tech2Wild built it within hours, on the smaller pack. His repo serves it at tensor-parallel 2 over RoCE. The weights sit on one node, and the other reads them over NFS.
| Two Sparks, 2.15 bpw, TP2 | Result |
|---|---|
| Context per request | 300,000 tokens |
| KV pool | 726,578 tokens |
| Single stream, code | 30.4 tok/s |
| Single stream, prose | 19.8 tok/s |
| Single stream, counting to 300 | 39.4 tok/s |
| Four streams, code, total | 66.2 tok/s |
| Mixed prompts, four streams, total | 46.5 tok/s |
| Cold prefill, 5K to 128K tokens | 501 to 567 tok/s |
| Smoke test | 4 of 4 |
Speculative decoding explains the spread. DSpark drafts three tokens a step, and its acceptance runs 71% on code but 33% on prose. Total throughput peaks at four streams. At five, per-stream speed halves, and the README says the cause is not yet investigated.
Prefill is the weak spot: a 128K-token prompt takes about four minutes to the first token. It took two small patches to the vllm-exl3 plugin to run at TP2 at all. The lane is text only, because the 2.15 bpw pack has no vision weights.
DeepSeek-1017 on two, three and four Sparks
Three Sparks, TensorFold against SGLang. @HandwerkAiLab ran DeepSeek-V4.1-Flash-1017 at TP3, with @easthambuilds' 3.5 bpw EXL3 quant on TensorFold.
| Three Sparks, TP3 | TensorFold | SGLang |
|---|---|---|
| Four sessions | 108.6 tok/s | 68.3 tok/s |
| Wait on a 16K agent prompt | 11.8 s | About 20 s |
He reports up to 242.9 tok/s across four streams, and 95.2% on his 133-task German DACH-Bench, or 97.5% with thinking on. He did not post the SGLang quant.
Two Sparks with an x86 helper. PLI Labs runs a 2.75 bpw EXL3 build of DeepSeek-1017 across two Sparks. A separate server with 512 GB of RAM holds the 189 GiB of Engram tables, and two NVLinked RTX 3090s run the drafter. The Sparks verify every token, so outputs keep the target model's distribution.
| Two Sparks + x86 server, one request | First token | Decode |
|---|---|---|
| 124-token prompt | 0.36 s | 49.7 to 53.1 tok/s |
| 32,000-token prompt | 14.5 s | 58.7 to 66.2 tok/s |
| 150,000-token prompt | 68.6 s | 53.0 to 54.9 tok/s |
Those runs use temperature 1 and thinking on at high effort. Three 32K requests at once made about 105 tok/s combined. Prompt reading runs about 2,200 tok/s.
Two Sparks, TensorFold TP2. @bertholomusai's v0.7.1 reads a 93K prompt at 2,355 tok/s, 10% faster than v0.7.0, with decode flat and replies byte-identical. @RipperdocMD runs it on his two Sparks at about 75 tok/s with 512K context.
Four Sparks. @sash__mit runs DeepSeek-1017 on four Sparks with a 3.5-bit EXL3 quant by @Libertai_DAI, about 334 GB, on TensorFold TP4. He reports 70 to 150 tok/s, about 3,000 tok/s prompt reading and 16 parallel streams, with no benchmark file attached.
Ling 3.1 Flash: four Sparks against an API
Friday's update reported @sudoingX's 12-hour run of Ling 3.1 Flash on four Sparks. He has now run the same prompt against an API with the same agent settings.
| Ling 3.1 Flash, one prompt | Q4 GGUF, four Sparks, llama.cpp RPC | API, Novita on OpenRouter |
|---|---|---|
| Wall time | 12h 26m | 57 min |
| Decode | 7.5 tok/s | 98.6 tok/s |
| Tokens out | 325k | 228k |
| Tests | 12 of 12 plus visual gate | 12 of 12 plus visual gate |
RPC splits the model by layers, so the four boxes take turns on every token. That caps it near 12 tok/s, he explains. Novita keeps its serving precision private, so he labels that side simply as the API. His conclusion: the speed gap is the setup, and an official 4-bit build would make the model practical locally.
The same model off the Sparks
- Mac Studio M5 Ultra, 256 GB. @niklaslenz_ai ran antirez's Q2 GGUF of DeepSeek-1017 in a fork of the ds4 engine at 54.5 tok/s, code and prose. The 189 GiB of Engram tables stream from SSD.
- One 12 GB GPU. Strata 0.2.4 runs ISTA-DASLab's 2-bit DeepSeek-1017 GGUFs on one 12 GB card with 46 to 64 GB of RAM. On one RTX 5090 with 128 GB of RAM, @jun1228909 reports about 14 tok/s with reasoning off.
- Qwen3.8-Flash-Next on a Mac. @ashxhart reports over 150 tok/s on prose and 230 on code on his M3U Mac Studio with an experimental kernel, up from about 110.
Still open
- Four Sparks on DeepSeek's official weights. @MiaAI_lab showed the build running on Friday but has posted no speeds. She says she is busy on DeepSeek.
- Eight Sparks. No first-hand speed report yet. The figures in circulation are still NVIDIA's own. One owner of eight Sparks says he is starting to regret not buying RTX 6000s instead, with no numbers.
- The 64 GB Spark. It went on sale Friday. We found no owner report yet, only restatements of NVIDIA's launch details.
Sources
- @ViC305's two MiniMax quants (sizes, agreement, KL divergence, one-Spark recipe status)
- tonyd2wild/MiniMax-M3.1-Flash-EXL3-TP2-2x-DGX-Spark (two-Spark recipe, results, patches, the 4.26 bpw fit) and @Tech2Wild's post
- @HandwerkAiLab's three-Spark comparison and his earlier reply
- PLI Labs progress report and @NicoTheGreco's post
- bertholomus/deepseek-v4.1-tensorfold-tp2-2xgb10, @bertholomusai's v0.7.1 post and @RipperdocMD's run
- @sash__mit's four-Spark post
- @sudoingX's Ling comparison, his turn analysis, his setup notes and the floating-tree spec
- @niklaslenz_ai's Mac Studio run
- Strata releases, the ISTA-DASLab model card and @jun1228909's run
- @ashxhart's Mac kernel numbers
- @MiaAI_lab on her DeepSeek work and an eight-Spark owner's reply