Sparks Update: DeepSeek-1017 on Two Sparks Against Four, and GLM-5.3-Flash Measured

One builder ran the same agent prompt on two and four Sparks, GLM-5.3-Flash got a full two-Spark workup, and MiniMax went under Kubernetes on four.

Saturday's Sparks Update ended with no four-Spark speeds for DeepSeek-1017 and the two-Spark MiniMax pack still waiting for a recipe. On Saturday and overnight, builders filled in part of that. Every number below names its build and links its source.

At a glance

Two Sparks against four on DeepSeek-1017. The same voxel-world prompt took 73.7 min on two Sparks and 49.7 min on four, with different quants.
GLM-5.3-Flash on two Sparks decodes code at 51.2 tok/s and serves 99.2 tok/s across eight requests on @MiaAI_lab's recipe.
Qwen3.8-Flash-Next on one Spark or two. NVIDIA's own vLLM command gives 94 tok/s on JSON at TP=2, against 62 on one Spark.
MiniMax-M3.1-Flash on four Sparks runs at 59.0 tok/s single-stream under Kubernetes. A two-Spark quant is done, and its recipe is in progress.
TensorFold 1.1.4 makes Qwen3.8-27B cold prefill on GB10 up to 5× faster.
Still missing: a same-quant comparison of two and four Sparks, and DeepSeek's official weights measured on four.

DeepSeek-1017 on two Sparks against four

@niklaslenz_ai ran @bertholomusai's TensorFold recipes for DeepSeek-V4.1-Flash-1017 on two Sparks and on four. Both builds got the same voxel-world prompt in DeepSeek Harness 0.2.1-rc.2, with thinking on and one try.

Measure	Two Sparks	Four Sparks
Quant and split	EXL3 2.5 bpw, TP2	EXL3 4.0 bpw, TP4
Recipe	v0.7.2 (commit 30a5bf9)	commit 8f84e7a
Agent time	73.7 min	49.7 min
Tokens written	138,675	121,707
Browser check	8.0/10	9.0/10 (older capture method)
Decode, code	70.5 tok/s	82.8 tok/s
Decode, prose	70.6 tok/s	80.8 tok/s
Eight requests at once, total	106.3 tok/s	141.2 tok/s

Both ran with 1,048,576 tokens of context, the recipe's setting. On two Sparks he also measured 1,830 tok/s reading a 32K-token prompt, and 17.9 s to the first token.

The two-Spark recipe wants RDMA GID index 5, which his Sparks did not have. He gave the link ports IPv4 addresses instead. The recipe runs 4 slots, so at eight requests the rest wait for a free slot.

GLM-5.3-Flash on two Sparks

@niklaslenz_ai also ran GLM-5.3-Flash on two Sparks with @MiaAI_lab's recipe at commit 3fd60dfc. It uses an EXL3 4 bpw quant at TP2, a DFlash2 draft model and the TensorFold Python engine.

GLM-5.3-Flash, two Sparks	Result
Decode, code	51.2 tok/s
Decode, prose	48.6 tok/s
Draft acceptance	61.7%
One request	46.6 tok/s
Four requests at once, total	80.2 tok/s
Eight requests at once, total	99.2 tok/s
Reading a 32K-token prompt	1,944 tok/s
Memory test, 131,072-token window	About 100 GiB per Spark

As an agent in OpenCode 1.19.3, it built a voxel sky kingdom in 71.1 min and scored 9/10 on his browser check. Mia's Zig-engine branch, which her docs call work in progress, decoded code at 26.6 tok/s and took 98.8 min.

@bertholomusai says GLM 5.3 Flash has been his daily driver on two Sparks for almost a month. He now runs it on four, while he experiments on another four.

Qwen3.8-Flash-Next on one Spark or two

@atasrb ran Qwen3.8-Flash-Next with NVIDIA's own vLLM cluster command, FP4-Local at TP=2. He reports one user, MTP 4 and the median of 5 runs.

Task	One Spark	Two Sparks
JSON	62 tok/s	94.1 tok/s
Code	58 tok/s	89.7 tok/s
Rewrite	Not posted	71.0 tok/s
Explain	Not posted	68.1 tok/s
Story	Not posted	50.6 tok/s

The server starts in about 10 minutes, with 262k context per request. He did not measure quality himself. He cites NVIDIA's model card: GPQA Diamond 91.5% for this FP4 build, against 92.2% for BF16.

MiniMax-M3.1-Flash on four Sparks, and two next

@NftHoss runs MiniMax-M3.1-Flash on four Sparks with SGLang at TP4. The Sparks talk over a patched NCCL ring, a switchless overlay across four RoCE HCAs per node. The weights are 250.1 GB in mxfp8.

MiniMax-M3.1-Flash, four Sparks	Result
Context	1,048,576 tokens
Single-stream decode	59.0 tok/s
Six streams, total	122.9 tok/s
Prefill at 32k	1,121 tok/s
Prefill at 266k	1,248 tok/s
Boot to first response	~12 min

He says vLLM's pynccl path could not start on his ring, while SGLang's could. Loading 250 GB of weights over 1 GbE NFS was too slow, so he fed them over the 200G fabric.

For two Sparks, @jayleaton says his MiniMax Flash quant is done and the recipe is in progress. He built it from a baseline recipe by @MiaAI_lab and is porting in the latest TensorFold. @JASONMCNAB reports a different two-Spark model, Red Snow 5.3 Flash FP4, at 26.8 tok/s on his Hermes Agent test.

What TensorFold 1.1.4 changes on GB10

@ashxhart's TensorFold 1.1.4 brings a new Qwen3.8-27B recipe on CUDA. On GB10, cold prefill improves by 54% at 2k, 107% at 8k, 258% at 32k and 416% at 64k.

GB10 result	Before	After
Qwen3.8-27B cold prefill at 64k	203 tok/s	1,045 tok/s
Gemma 4 12B chat with its MTP assistant	30 tok/s	67 tok/s
Gemma 4 31B with z-lab's DFlash drafter	12.6 tok/s	29.9 tok/s

Flash Next serving gains 35%, 75% and 116% at 2, 4 and 8 streams, with tokens unchanged. Release binaries do not serve DeepSeek checkpoints yet. @MiaAI_lab says she will update any of her recipes that can benefit.

Off the Sparks

Three engines on one RTX 5090. @niklaslenz_ai ran Qwen3.8-27B with the same agent prompt on each.

Engine	Code decode	8 at once, total	Build time
@BobClawblaw's TensorFold recipe + DFlash2	268.2 tok/s	204.6 tok/s	31.1 min
Neroued's NInfer, NVFP4 + MTP3	169.0 tok/s	711.9 tok/s	56.5 min
@MiaAI_lab's kit, EXL3 4.0 bpw + MTP	154.5 tok/s	One request at a time	61.6 min

Swipe the table sideways to see every column.

Four CMP 170HX cards. @Morrowmake says four 170HXs cost the same as two Sparks and run DeepSeek 2-4x faster. His recipe reports 287 t/s on code for one user, and 3800t/s cold prefill.

A Mac. @jundotkim's oMLX 0.7.3.dev1 runs MiniMax M3.1 Flash on an M3 Ultra at 42.0 tok/s, up from 22.5, with its bundled DSpark draft.

One consumer GPU and RAM. @Goldstein1337 runs GLM-5.3-Flash on a 3090 Ti with 64GB of RAM at 20 tok/s, with Project Maya. @Gould0100 runs DeepSeek V4.1 Flash with Strata 0.2.4 on an RTX 5090 at about 30 tok/s, at a 256K context setting.

Still open

A two-against-four Spark comparison on the same quant and recipe.
DeepSeek's official weights measured on four Sparks.
Speeds for @jayleaton's two-Spark MiniMax recipe.
@mmastrac's remote KV for gr8, which would serve long GLM 5.3 sessions from spare RAM on other Sparks, has no published numbers yet.