DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-1M-57toks
DeepSeek-V4-Flash-DSpark abliterated (uncensored) · ~100% refusal bypass · C1 ~57 tok/s · 1M ctx · 2× DGX Spark · HF weights
- 59
- Python
- 3 mo ago
drowzeys
Built from drowzeys's public GitHub profile. Not on Take My Repo yet, so there are no posts or updates here.
DeepSeek-V4-Flash-DSpark abliterated (uncensored) · ~100% refusal bypass · C1 ~57 tok/s · 1M ctx · 2× DGX Spark · HF weights
MiniMax-H3 on single DGX Spark: keys Sol-Engine ports + Heretic TE + Spectrum/Sage speed + 0.5MP upscaler finish. Formal bench included.
1.0 Beta · oMLX 0.6.3rc2 Dual-ANE+CPU · Qwen3.8-27B abliterated oQ4e · Mac Studio M3 Ultra
keys-DGX-Sparkticus Utimate Power Pack Unleashed (Dual DGX-Sparks Needed)
MiniMax-H3 33B audio+video DiT accelerated 1.54x on a single NVIDIA DGX Spark (GB10, sm_121): NVIDIA Sol-Engine FirstBlockCache + batched VAE decode ported to single-GPU ComfyUI, Sol-Attn patched for compute 12.1, kijai Triton INT8/TMA kernels. Benchmarks, charts, one-shot installer, agent recipe.
MiniMax H3 (video + audio) on TensorFold 0.6.5 with an H3 family, Mac Studio M5 Ultra: 5 s 864x480 clip in ~50 s, 40 s at 768x448 — the fastest H3 we have measured on Apple Silicon; int8 tensor-unit kernels + Turbo adapter; one-shot + GHCR carrier
GLM-5.3-Flash (Z.ai) NVFP4 (LibertAI) + NVFP4 KV, 1M context on 2x DGX Spark GB10 — vLLM TP2 recipe, walls ledger, benchmarks, needle to 1.04M; abliteration with blackforest_AI next
English guide + one-command patcher for the vLLM spin-wait bug that overheats NVIDIA DGX Spark / GB10. Credit: nacyot. CPU cores spin at full clock waiting on shm_broadcast; 1s -> 2ms drops SoC up to 24C with no throughput cost.
Autonomous self-improving 4x DGX Spark (GB10) MoA stack + LoRA loop (DSV4F router, Qwen3.6/Omni/TwoTower/Gemma). Hermes MoA routing, ~90% local, self-LoRA. Includes a SINGLE-SPARK recipe (all-Gemma-4: 31B brain + DiffusionGemma + 12B) to build the same on 1 GB10.
Qwen3.8-27B ADay777-Ablit · NVFP4 · A4Q · NVFP4-KV (4M+ pool) · MTP-3 — single DGX Spark (GB10/sm121), eugr vLLM 0.27. 1M ctx @ c=4.
Full GLM-5.3 (753B) served by TensorFold with tensor parallel 4 across four NVIDIA DGX Sparks: results, recipe, node settings, benches
Run on TWO-DGX-Spark - vLLm-0.24.0 dual cache optimized DSV4F+DSpark+NVFP4 KV (Concurrency 12 with 1.5M context/3M KV token Pool) >0.58-0.6 acceptance.
Boo, a ghost-faced audio/visual agent you talk to from your phone: speech in and out, camera, hand tracking, memory and a tool-using back agent — powered by TensorFold, on DGX Spark and Apple silicon
keys-latest: GLM-5.2 QuantTrio INT4/INT8 ablit + DFlash · 4x DGX Spark · k=12 · 128k · seqs=4 · KV 10GiB
GLM-5.2 Quantrio INT4/INT8 Mixed Abliterated — SPEED=1 C1≈30 tok/s @ 128k on 4x DGX Spark. Step-by-step recipe, image bake, results.
Concurrency Patch to enable on DSpark - DeepSeek V4 Flash/Pro C=16 can be modified for more
DeepSeek-V4.1-Flash TR3-Hybrid: 1M context, 8M+ KV pool (9,452,923 tokens) on 4x DGX Spark, TP4, CUDA graphs + DSpark k=5 + vision + tools. Pool grown via GMU 0.83 + batch 2048. With abliterated option.
keys- 1M context: Inkling-Small-NVFP4 + DSpark + NVFP4 KV on SGLang, sm_121a-optimized on Two DGX Sparks — one-shot agent-executable recipe, all GB10 patches + bench numbers
GLM-5.3-Flash Abliterated oQ4-MTP on oMLX 0.7.0, Mac Studio M5 Ultra: 60.9 tok/s greedy decode (2.6x the M3U pack), ~1.1k tok/s prefill, Lightning MTP fixed depth 2, one-shot setup
Stock MOPD C3 (hybrid-21) Jarrelscy ARVQ on 4x DGX Spark, 1M context. Prebuilt v4 image.
MiniMax Code CLI enhancement pack: continuous 25% context renewal, context progress meter, browser tooling, Syntra and djev-spark local decision integrations.
DeepSeek-V4.1-Flash 763B on ONE 256 GB Mac Studio M3 Ultra - oMLX 0.7.0.dev2 + CED prefill + 14-layer 2-bit requant + SSD-paged experts. 29 tok/s code, 551 tok/s prefill.