Reproducible distributed LLM inference, quantization research, and evidence-backed deployment recipes for NVIDIA GB10 / DGX Spark systems.
We publish the receipts that matter: pinned revisions, integrity manifests, exact runtime formulas, bounded capability gates, concurrency/context evidence, restart proof, limitations, licensing, and upstream credits.
| Project | Topology | Focus | GitHub | Hugging Face |
|---|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 on TensorFold (Zig) | 4× GB10, TP4 | Our 4.0 bpw quant; exact 2-D tensor-parallel split; 1M tokens per request, 16 streams on a 20M-position shared window; vision, structured output, DSML tool calls, DSpark drafting | Recipe + evidence | Card · Weights |
| DeepSeek-V4.1-Flash-1017 on TensorFold (Zig) | 2× GB10, TP2 | Our 2.5 bpw quant sized for two nodes; clean-room dsv41 family on TensorFold 1.0.12's native engine; 1M window, vision, DSpark drafting |
Recipe + evidence | Card · Weights |
| GLM-5.3 (full, 753B) EXL3 3.0 bpw | 4× GB10, TP4 | Our own quant; TensorFold engine branch; 1M context via context parallelism; MTP drafting; KL 0.109 vs BF16 | Recipe + benchmarks | Weights |
| DeepSeek-V4.1-Flash-1017 configuration | 2× GB10, TP2 | Verified .env and measured rows for the 1017 update on a third-party 2.9 bpw quant (LibertAIDAI, MIT); GB10 memory notes |
Recipe | — |
| DeepSeek-V4.1-Flash DSpark | 4× GB10, TP4 | SGLang recipe v1.1.6: measured deviations from upstream, 1M context, 8M-token shared KV pool, systemd lane | Recipe + evidence | Card |
| GLM-5.3-Flash NVFP4 | 4× GB10, TP4 | 1M-configured context, FP8 KV, DFlash2, tools, multimodal and long-context gates | Recipe + evidence | Evidence card |
| DeepSeek V4 Flash Graph-8 | 4× GB10, TP4 | Four-node scaling, evidence and release recipe | Repository | Model card |
| DeepSeek V4 Flash Graph-8 | 2× GB10 | Reproducible two-node deployment and benchmarks | Repository | Model card |
- GLM-5.3-EXL3-3.0bpw: full GLM-5.3 for four GB10 nodes.
- DeepSeek-V4.1-Flash-1017-EXL3-4.0bpw: for four GB10 nodes.
- DeepSeek-V4.1-Flash-1017-EXL3-2.5bpw: sized for two GB10 nodes.
- clef-quant (in development): low-bit Cloudflare Clef 27B for one 24 GB GPU, with calibration-first gates. Nothing published yet.
- TensorFold (fork of ashhart/TensorFold): our DeepSeek-V4.1-Flash family on TensorFold 1.0.12's native Zig engine, branches
deepseek-v41-zig-tp4(four Sparks) anddeepseek-v41-zig-tp2(two Sparks); branchglm-dsa-tp4runs the full GLM-5.3 over four Sparks. The original Pythondsv41engine stays ondeepseek-v41-tp2. Community PRs welcome (thanks @Blackwellboy for #33 and #34). - MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks #10: per-node active InfiniBand HCA detection for TP4 boot on mixed-port fleets.
- NousResearch/hermes-agent #95120: durable Discord thread rename and slash-command sync.
- dsh-public-plugins: community plugins for DeepSeek Harness (DSH) Desktop: tailnet GUI access, Discord threads as sessions, coordinator threads, calendars, session bridge API, autopilot, decisions inbox, workspace categories and more.
- Attribute upstream models, checkpoints, runtimes, kernels, and recipe lineage.
- Separate configured capability from directly proven capability.
- Publish machine-readable evidence and known limitations alongside headline results.
- Never present third-party quantizations as BertholomusAI-created artifacts.
- Mirror weights only when provenance and redistribution rights are clear.
- Prefer exact revisions and hashes over mutable tags.
- DeepSeek-V4.1-Flash-1017 on TensorFold: faster long-prompt prefill and more concurrent streams on two and four GB10 nodes.
- Full GLM-5.3 on four GB10 systems: faster prefill and vision for the non-Flash model.
- New model families and multi-node tensor parallelism for TensorFold on GB10.
- Full-download quantization workflows with explicit authorship and reproducibility.
- Speculative decoding, long-context validation, and safe production supervision.
- Hermes-native local-agent infrastructure.
Hugging Face: huggingface.co/bertholomus · X: @bertholomusai