Sparks Update: Eight-Node Clusters, MiniMax-M3.1-Flash on Day One and a 64 GB Spark
NVIDIA made eight-Spark clusters official, four Sparks ran MiniMax’s new 428B model within hours, and two-Spark builds wait on their first numbers.
Yesterday's State of the Sparks covered two and three DGX Sparks. Overnight the ceiling moved. NVIDIA now supports eight-node clusters, MiniMax published a 428B model that four Sparks ran on day one, and a cheaper 64 GB Spark goes on sale today. Here is what changed, with every number tied to its build.
At a glance
- Eight Sparks are official. NVIDIA supports eight-node clusters with 1 TB of pooled memory. It measured the full GLM-5.3 at 31 tok/s for one user on eight boxes.
- MiniMax-M3.1-Flash runs on four Sparks. A day-one recipe on MiniMax's own weights measures 46.5 tok/s on prose at one stream, and 73 of 75 on a quality check.
- Two Sparks: files exist, results do not. A 3.06 bpw EXL3 pack and a 203.5 GB GGUF fit two Sparks by size. Neither has a published speed or quality number on Sparks yet.
- DeepSeek-1017 gets an experimental quant. Mia's EAQ v1.1 keeps the sparse-attention indexer at source precision. She says EAQ v1 stays the one to use.
- A 64 GB Spark at $4,999 ships from six partners today. Two of them pool 128 GB.
- Four Sparks over llama.cpp RPC is slow. @sudoingX ran the 560B Ling 3.1 Flash at 7.5 tok/s, and it still passed its own tests after 12 hours.
NVIDIA makes eight Sparks official
NVIDIA announced official eight-node DGX Spark clusters at GTC Berlin on October 22. One Spark has 128 GB, four reach 512 GB and eight reach 1 TB. Setup is manual for now. NVIDIA Sync's Cluster Assistant gains eight-node support by the end of the year, the company says.
The headline test is the full 753B GLM-5.3, which needs more than four boxes. NVIDIA ran its own NVFP4 checkpoint, about 470 GB, with MTP3 speculative decoding:
| Users at once | Prefill, total tok/s | Decode, total tok/s | Decode per user, tok/s |
|---|---|---|---|
| 1 | 1,600 | 31 | 31 |
| 4 | 1,600 | 51 | 21 |
| 8 | 1,600 | 67 | 18 |
NVIDIA also ran GLM-5.3-Flash across eight nodes at roughly 80 tok/s in generation. It did not publish the quant or prompt for that figure.
To isolate the network, NVIDIA timed a dense Qwen3.8-27B in NVFP4 with no speculative decoding, at 8,192 tokens in and 256 out. Decode climbs from 11.86 tok/s on one Spark to 52.58 on eight.
Two wiring choices matter. A four-box ring with no switch is the cheapest four-node setup, and NVIDIA calls its throughput impressive for this workload. Eight boxes, split clusters and its new RoCEnante collective need a switch. RoCEnante is a one-shot all-reduce merged into FlashInfer. It falls back to NCCL if its startup test fails, and NCCL stays the default as clusters and concurrency grow.
The 64 GB Spark
DGX Spark 64GB goes on sale Friday, October 23, starting at $4,999, from Acer, ASUS, Dell, Gigabyte, HP and MSI. It keeps the GB10 chip, ConnectX-7 networking and DGX OS. NVIDIA says one runs models up to 100 billion parameters, and two cable together for 128 GB. Tom's Hardware puts a 128 GB GB10 system at roughly $7,000 to $9,000 right now, when you can find one.
For the models in this feature, a 64 GB box changes little. Every build below needs 145 GiB or more.
MiniMax-M3.1-Flash, checked
MiniMax published the weights on Thursday evening, Pacific time. The model card lists about 428B total parameters and 23B active per token, a context of up to 1M tokens, and text, image and video inputs. Sparse attention runs in every layer. The experts and attention were trained with quantization, and the release includes a DSpark draft head for speculative decoding. The checkpoint is 252.4 GB on disk, per the four-Spark recipe below.
MiniMax's own table includes DeepSeek's newest Flash, and DeepSeek wins most coding rows:
| Benchmark | MiniMax-M3.1-Flash | DeepSeek-V4.1-Flash-1017 |
|---|---|---|
| DeepSWE v1.1 | 73.7 | 77.0 |
| Terminal Bench 4.0 | 38.1 | 42.3 |
| NL2Repo-Bench | 61.2 | 68.2 |
| Automation-Bench v1.0.6 | 62.8 | 60.5 |
| HLE (without tools) | 43.4 | 42.6 |
| BabyVision (with tools) | 91.0 | 92.5 |
These are vendor numbers for the official weights. The card says third-party quantizations may differ. The license is MiniMax's own: @superalesha reads it as free for commercial use with a "Built with MiniMax" credit, with permission needed above $20M in yearly revenue.
What fits
| Build | Size | Fits | Speed on Sparks | Quality check |
|---|---|---|---|---|
| Official NVFP4 experts + MXFP8, four Sparks, TP4 | 252.4 GB, about 63 GB a box | Four Sparks, 262K context | 46.5 tok/s prose, one stream | qeval 73 of 75 |
| AtomicChat GGUF Q4_K_M | 258.0 GB, 240.3 GiB in memory | 384 GB machines and up | Not published | Pending |
| AtomicChat GGUF Q3_K_M | 203.5 GB, 189.5 GiB | 256 GB machines | Not published | Pending |
| Mia EXL3 EAQ v1, 3.06 bpw | About 171 GB repo, 167 GB text model | Two Sparks, by size | Not tested | Not tested |
| AtomicChat GGUF Q2_K | 155.9 GB, 145.2 GiB | 192 GB machines | Not published | Pending |
AtomicChat says its release-day GGUFs are plain quants without an importance matrix. Its smaller Atomic Dynamic files and quality numbers come later. In llama.cpp the attention cache takes about 0.14 MB per token, or 4.5 GB at 32K. Sparse attention runs only with flash attention on (-fa on) and without --kv-unified; otherwise llama.cpp falls back to dense attention.
Mia's 3.06 bpw pack stores every routed expert at 3 bits. It keeps the sparse-attention indexer at FP16 and the router at FP32. Her card is plain about the gaps: no quality number, no long-context test, no speculative-decoding test, and no engine yet verified to load it.
The four-Spark build
@majewskizby published the first measured bring-up, knapcio/MiniMax-M3.1-Flash-4x-DGX-Spark-TP4, on four Sparks through a MikroTik CRS812 switch. His first post gave 39.7 tok/s prose and 62.2 structured at one stream. By Friday morning the repo's default had moved to SGLang with RadixArk's DFlash2 draft and RoCEnante collectives:
| Test, default profile (SGLang) | Result |
|---|---|
| speedbench, 73 prompts, one stream | 45.5 tok/s geometric mean (prose 38.7, long context 56.8) |
| sparkDash prose, one stream | 46.5 tok/s |
| sparkDash structured, one stream | 85.0 tok/s |
| sparkDash prose, 16 streams | 231.6 tok/s total, 15.3 per stream |
| Cold prefill, 4K to 128K | 1,421 to 1,490 tok/s |
| qeval, 75 tasks | 73 of 75 |
| Needle retrieval | Pass at 116K and 218K tokens |
The vLLM profile with MiniMax's bundled DSpark draft is the alternative. It prefilled a 58K-token prompt at 2,171 tok/s against 1,442 for SGLang, and boots in about 5 to 6 minutes instead of about 12. Every lever in the repo is lossless, it says: no requantization and no training.
GB10 is sm_121, so several of MiniMax's defaults do not run. The NVFP4 KV cache and MiniMax's sparse-attention kernels are SM100-only. The recipe uses FP8 KV cache, Triton sparse attention and Marlin kernels for the experts.
For scale, @Dataphil3 logged 212.6 tok/s at batch 1 on four RTX PRO 6000 Blackwell Max-Q cards with 96 GB each, on vLLM TP4, verified on LocalMaxxing.
DeepSeek-1017: a sharper indexer, and a claim
Mia published DeepSeek-V4.1-Flash-1017 EXL3 EAQ v1.1, marked experimental. It has the same 2.9 bpw body as v1. The change is the sparse-attention indexer path, which v1.1 stores in BF16 at the source values.
| Check | EAQ v1 | EAQ v1.1 |
|---|---|---|
| Same token picks as the original, fed the original's inputs, past 16K | 94.95% | 403,560 of 403,560 |
| Picks in a full quantized run (recall@512) | About 0.73 | About 0.73 |
| Size | 210.58 GB of tensors | About 0.15 GB more |
The second row is the honest part. In a full quantized run, the inputs reaching the indexer have already shifted, so a BF16 indexer alone cannot close that gap, Mia writes. Speed and generation quality are untested, and she says she has no spare compute to test it yet. The Engram tables are still separate: you need the base model's shards 47 and 48, about 95 GiB each.
Mia also says DeepSeek-1017 on TensorFold for four Sparks, on the official weights, is "in the works."
Four Sparks, one prompt, twelve hours
@sudoingX borrowed four Sparks from @niklaslenz_ai to run Ling 3.1 Flash, a 560B MoE that does not fit on his two. There was no official FP4 build for vLLM yet, so he used AtomicChat's AD-Q4_K_M GGUF over llama.cpp RPC, with Hermes Agent driving.
| Ling 3.1 Flash, Q4 GGUF, four Sparks, llama.cpp RPC | Result |
|---|---|
| Wall time on one prompt | 12 hours 26 minutes |
| Requests and output | 86 requests, 325K tokens |
| Decode | 7.5 tok/s ("painful," he writes) |
| Peak context | 143K of 262K |
| Tests | 12 of 12 pass |
| Code | 2,166 lines across 12 modules |
Four of its turns took about 9 hours and 77% of all tokens while writing very little. Next he runs the same prompt against the full model on the API, to see whether those long turns come from the model or his setup.
Engines
- TensorRT-LLM 1.4.0rc1 (October 21) adds DeepSeek-V4.1-Flash support and NVFP4 GLM-5.3-Flash checkpoints, and keeps DeepSeek-V4 tensor-parallel decode all-reduces off NCCL. Its notes name no Spark recipe for either model.
- vLLM and SGLang shipped day-one MiniMax images. The four-Spark recipe runs the arm64 builds
vllm/vllm-openai:minimax-m31-flashandlmsysorg/sglang:dev-minimax-m31-flash. - llama.cpp has had MiniMax-M3 support since July. AtomicChat checked its files with build
f5027fb21307on October 23. - TensorFold: @ashxhart posted “MiniMax M3.1 Flash with zero day support on TensorFold.” No Spark numbers yet.
The same level elsewhere
- AMD, DeepSeek-1017. @jgattermayer runs the Maya-S GGUF on a Strix Halo box with two RX 7900 XTX cards, streaming experts from an NVMe drive over USB4. He measured 37.1 tok/s after a cold 130K prompt (42.8 warm), and 31 to 35 tok/s across a 256K context ladder.
- Small models get faster without changing. In an open challenge on Macs and DGX Spark, Bonsai 2 9B decode went from 144.8 to 959.6 tok/s on Apple Silicon in a week. Every run had to produce exactly the reference engine's tokens, @pratikg explains. The record changed hands 158 times.
- One Spark. @MiaAI_lab's pick for a single 128 GB Spark is still Qwen3.8-Flash-Next. "That's 90% of what most users need," she writes.
Sources
- NVIDIA Technical Blog: Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes (Oct 22: eight-node support, GLM-5.3 and Qwen3.8-27B tables, RoCEnante, ring and switch)
- NVIDIA Blog: DGX Spark 64GB (price, partners, on-sale date, 100B parameters)
- Tom's Hardware on the 64GB DGX Spark (128 GB system prices)
- MiniMaxAI/MiniMax-M3.1-Flash model card (parameters, context, architecture, benchmark table)
- knapcio/MiniMax-M3.1-Flash-4x-DGX-Spark-TP4 (four-Spark recipe, speeds, qeval, GB10 changes) and @majewskizby's day-one post
- AtomicChat/MiniMax-M3.1-Flash-GGUF (file sizes, memory, llama.cpp notes) and AtomicChat's post
- Mia-AiLab/MiniMax-M3.1-Flash-EXL3-EAQ-v1-3.06-bpw (pack contents, untested items)
- @MiaAI_lab on the MiniMax weights, the two-Spark recipe and her later reply
- @superalesha on the MiniMax license, @ViC305 on a one-Spark quant, @ashxhart on TensorFold support
- @Dataphil3's four-GPU run and its LocalMaxxing entry
- Mia-AiLab/DeepSeek-V4.1-Flash-1017-EXL3-EAQ-v1.1-2.9-bpw, @MiaAI_lab's v1.1 post, her testing reply and the four-Spark DeepSeek note
- @sudoingX's 147.6 tok/s post
- @sudoingX's Ling 3.1 Flash run, his follow-up and the quant he used
- TensorRT-LLM v1.4.0rc1 release notes
- @jgattermayer's Strix Halo report and the Project Maya issue
- @pratikg on the Bonsai 2 9B challenge
- @MiaAI_lab's one-Spark pick