On October 16, a community recipe put the full GLM-5.3 on four DGX Sparks, at tensor parallelism 4. The model is GLM-5.3 (753B), and the four NVIDIA DGX Spark units run it. Underneath is gr8, a from-scratch inference stack for the GB10 processor. The same weekend, keys/Wasserman AI Studio announced release 2.8.0 for Mac and for Linux on DGX Spark (release post). It is a media-production suite built on the open-source TensorFold engine. The more consequential of the two, however, is the engine. An independent team writing a serving stack from scratch is the strongest evidence a hardware tier has reached platform status.
by Chris DePuy, 650 Group / October 28, 2026
gr8 is not a fork of an existing engine. The release post calls it a from-scratch engine for GB10 with its own RoCE fabric. In addition, it adds batched speculative decoding and per-round drafter selection. It reports the full model already running at 29 tokens per second out of the box (post). The engine lives at kindlingai/gr8. The recipe pins the engine to upstream commit c53b5b8 rather than forking it. It ships no weights, only patches, a build, an example configuration and measured results (recipe README). The model itself is z.ai’s open 753B release on Hugging Face. The 753-billion figure is the release post’s label, not a claim read from the model card.
How fast is GLM-5.3 on four DGX Sparks in practice?
It depends on what the configuration must do. Single-stream decode is the interactive user’s number. The aggregate columns are the multi-tenant operator’s. In addition, cold prefill decides how long a long prompt waits before its first token.
As of October 2026, the README’s speed tables, measured on October 19 on the v0.9 configuration, put single-stream decode at 40.6 tokens per second for prose, 60.6 for code and 75.5 for structured output. At the same time, the recipe was at release v0.10 as of October 25. Specifically, the runs used temperature 0 with thinking on and effort low. The three full runs behind those medians agree within 0.1 tok/s (README). Aggregate end-to-end throughput reads 47.0 tokens per second at one stream, 59.9 at two and 76.6 at four. Meanwhile, the per-stream values fall from 52.4 to 23.4 as sessions share verify steps. For context, the four-Spark NVFP4 receipts cover official NVIDIA checkpoints on the same hardware.
Cold prefill holds 1,381 tokens per second on an 8K prompt, and still 1,226 at 128K. The needle test at 96K tokens returned the correct answer (README). For example, a second suite, speedbench 0.2, ran a frozen set of 73 prompts across six categories. It measured the full GLM-5.3 at a geometric mean of 49.4 tok/s, with a median first-token latency of 175 ms.
For contrast, the release post’s figures at recipe v0.5 on October 16 read 38.0 prose, 59.1 code and 70.2 structured. Specifically, the runs used one stream. The two-week gap between the announcement table and today’s table is itself the release cadence of this project (post).
The GLM-5.3 weights: EXLR8 quantization and a DFlash2 drafter
The weights complete the stack: kindlingai’s EXLR8 quantization (width manifest home) and the target model’s own MTP layer. In addition, they pair with the incoai DFlash2 drafter (weights, drafter). The the Flash-variant DFlash2 follow-up covers the drafter in context, and the DFlash2 speculative-decoding results give its speed case.
The GLM-5.3 patch series: thirty-one levers, mostly bit-identical
What turns the upstream engine into the serving configuration is a patch series. The release post counted 26 small patches at announcement, every lever gated behind its own off-by-default flag. Together they retune the decode attention kernels and pace L2 prefetching of upcoming weights into DRAM-idle windows. Specifically, they repair GPU TLB exhaustion in the MoE kernel at working sets around 74 GB (post).
The README, read on October 28, lists thirty-one small patches, three of which are backports. They backport upstream’s own open pull request. In addition, every patch sits behind an off-by-default flag. Many levers from earlier releases are now upstream code. The pinned commit for the full GLM-5.3 recipe carries the fabric MTU fix, the GEMV splits and the INT4 draft heads. It also carries the route overlap and three rounds of decode attention kernels. Specifically, it merged as upstream pull requests #82 through #90 (README). The the 24-second-lesson upstream report covers the same upstream gap.
In practice, the MTU fix is the one that matters on a switched network. Without it the engine does not boot on a switched fabric. See the open RDMA bridge work for the fabric path. As the README puts it, “negotiate the RDMA queue-pair MTU from both ports’ active MTU instead of assuming one” (README).
GLM-5.3 quality: checked, not assumed
The README classifies the bit-exact levers with dedicated benches (l2pf_bench, k1_audit). It records the same T=0 top-20 log probabilities on the cluster’s 75 qeval prompts, 75 of 75 with KL 0. Specifically, each reading was taken on a freshly booted cluster (README). For the GLM-5.3 release line, the v0.9 configuration against dh32f shows a mean KL divergence of 2.830e-3 over 3,622 shared positions. One configuration run at concurrency 1, another at 4, differing by roughly 5.0e-3 for scale. In addition, the post states that pull requests upstream are in flight. As a result, the patches ride the engine’s mainline rather than a private fork (post).
keys/Wasserman AI Studio: a full media pipeline on the same box
The second October 16 release for the GLM-5.3 hardware runs in the other direction: not a lower layer, but a complete application. The post announces keys/Wasserman AI Studio at version 2.8.0, powered by TensorFold, for Mac and for DGX Spark under Linux.
For instance, its item list runs from a MiniMax H3 implementation and Qwen Image 2.1 Turbo. The list reaches Ref2VA, Fish Audio Drama3, Flux-3 Klein and Motion Previs. The release also adds custom Metal and CUDA cores, a hybrid ComfyUI workflow and an integrated prompt and scene producer. It brings parallel processing across DGX Sparks and a 2x 2K upscaler built in (release post).
In a follow-up reply the same author states that 30-minute videos are within scope (length reply). A second reply links the produced results in the repository (results reply, repository). For instance, TensorFold is Ash Hart’s open-source inference engine for the Metal, CUDA and Vulkan backends (project). See our earlier TensorFold decode-speed work for the engine’s numbers.
What the two releases amount to
A from-scratch engine and a complete application appeared in the same week on the same class of hardware. That is what a platform tier looks like once independent teams stop waiting on the vendor’s own stack.
However, no second team has replicated any figure yet. Every number above is a single team’s measurement, published with the recipe, its patches and its measurement tools. As a result, it remains the operator’s own report until wider replication.
A purchase decision should size on the published tables for the full GLM-5.3 model on Spark clusters. That holds only after a second operator reads the same numbers on the same hardware. In addition, firms such as 650 Group have been tracking this shift in their AI infrastructure research. The question this quarter is how far GLM-5.3 on four DGX Sparks has moved from peripheral to core. For the platform-tier contrast, see what one Spark runs today.