Big open models on hardware you can own. We serve DeepSeek, GLM, Mistral, Beam, Ling, Qwen and others on NVIDIA DGX Spark (GB10) clusters and on single RTX PRO 6000 cards. We write our own engine families, make our own quants and publish every recipe with its measurements.
Each recipe pins its exact revisions and hashes, and ships its gate results, its limits and its credits.
- Qwen3.8-Flash-Next NVFP4 on one RTX PRO 6000 (v1.1): GPTQ-calibrated NVFP4 experts on the stock TensorFold 1.2.9 release, no fork and no patches. The full 262K context; 159 tok/s on one stream, 445 tok/s total with 8 agents.
- spark-attic (v0.1.7): about 2 GiB more memory for CUDA on a
headless GB10. The driver reserves memory for a display that a headless node never drives.
atticdlends it to the inference engine, so the KV cache grows with no other change. Replies stay identical and speed stays within noise in vLLM, SGLang and TensorFold. Upstream: TensorFold #2762, #2763. - GLM-5.3 (full) EXLR8 on four DGX Sparks (v2): the gr8 engine and exlr8 weights by KindlingAI (Matt Mastracci, mmastrac, and Chuck, chuck-ads / adapt-ai-systems). Our recipe adds a concurrency dial: 4 chats at once at 95 tok/s total (about 24 each), or one chat with up to 864K tokens (with spark-attic on). Our busy guard is merged upstream in gr8 #516: a request that would evict a live session gets a 429 at once, so agents do not go quiet.
- DeepSeek-V4.1-Flash-1017 on 4× GB10 (v0.3.7): 32 users at once, 290 tok/s total, first token in 0.42 s (was 15.9 s). One stream: 138 tok/s on code.
- DeepSeek-V4.1-Flash-1017 on 2× GB10 (v0.7.5): 32 users at once, 249 tok/s total (was 198). One stream: 114 tok/s on code. 1M-token prompts.
| Model | Hardware | Highlights | Recipe | Weights | Support |
|---|---|---|---|---|---|
| GLM-5.3 (full) EXLR8 | 4× GB10 | KindlingAI's gr8 + exlr8; concurrency dial, busy guard, optional spark-attic, SSD KV store | GitHub | KindlingAI | Active |
| Qwen3.8-Flash-Next | 1× RTX PRO 6000 | NVFP4 (GPTQ experts) on the stock TensorFold release; 262K context; 445 tok/s total with 8 agents | GitHub | NVFP4 | Maintained |
| Mistral Large 4 (1T) | 4× GB10 | Our mistral_large4 TensorFold family; the full 1M window on one stream; Mistral's draft head: ~30 tok/s on code |
GitHub | 2.5 bpw | Maintained |
| Beam-1.0 (501B) | 2× GB10 | Our TensorFold beam family; 868K-token window; 6–36 % faster than our vLLM lane on six of seven decode rows |
GitHub | 3 bpw | Maintained |
| Ling-3.1-flash-1019 (560B) | 2× GB10 | Our TensorFold bailing_v3 family; native 262K window; MTP drafting |
GitHub | 3.03 bpw | Maintained |
| Ripple-65B-Preview | 1× RTX PRO 6000 | NVFP4 + calibrated FP8 attention; 218 tok/s one stream on code | GitHub | NVFP4/FP8 | Maintained |
| DeepSeek-V4.1-Flash-1017 | 4× GB10 | Our TensorFold (Zig) family, exact 4-way split; 1M tokens a request; vision, tool calls, structured output, DSpark drafting | GitHub | 4.0 bpw | Maintenance |
| DeepSeek-V4.1-Flash-1017 | 2× GB10 | The same family on two nodes; 1M window, vision, DSpark drafting | GitHub | 2.5 bpw | Maintenance |
| GLM-5.3 (full) | 4× GB10 | Our TensorFold branch; 1M context through context parallelism; KL 0.109 vs BF16 | GitHub | 3.0 bpw | Frozen |
- TensorFold fork of ashhart/TensorFold
by Ash Hart (ashhart). One branch per model family:
deepseek-v41-zig-tp4,deepseek-v41-zig-tp2,mistral-large-4-tp4,beam-1.0-tp2,ling-3.1-tp2,glm-dsa-tp4. Thanks to Blackwellboy for #33 and #34. - kindlingai/gr8 #516 (merged): the busy guard.
- ashhart/TensorFold #2762, #2763 (open): the display carveout as a memory source for TensorFold.
- MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks #10 (open): per-node InfiniBand HCA detection for TP4 boot.
- dsh-public-plugins: plugins for DeepSeek Harness (DSH) Desktop: Discord threads as sessions, coordinator threads, calendars, a session API, autopilot, a decisions inbox and more.
- infiniquarium-wallpaper: a live, self-running aquarium as your macOS wallpaper, from the Infiniquarium trial.
Models and engines move fast, so each repository has one support level:
- Active: spark-attic, the GLM-5.3 EXLR8 recipe and the TensorFold fork. We triage issues and review pull requests every week, and we update them as the engines move.
- Maintained: we publish a new verified tag when an engine or model change needs one, and we answer issues.
- Maintenance: stable as pinned. We answer issues and review bug-fix pull requests. No new features.
- Frozen: the verified snapshot as pinned. No new support.
- Archived: read-only. The README links to the replacement.
A bug report helps most with your hardware, driver version, recipe tag or engine commit, and the logs.
Earlier recipes (archived)
- DeepSeek-V4.1-Flash DSpark, 4× GB10, SGLang
- DeepSeek-V4.1-Flash-1017 configuration, 2× GB10 (third-party 2.9 bpw quant by LibertAIDAI)
- GLM-5.3-Flash NVFP4, 4× GB10
- DeepSeek V4 Flash Graph-8, 4× GB10 and 2× GB10
- Every number is our own measurement on the stated hardware, or is marked as someone else's.
- We credit the model, the weights, the engine and the recipe each build stands on.
- We never present a third-party quant as ours.
- We pin exact revisions and hashes, and we list what is not measured yet.


