keys-TensorFold Studio (Mac+DGXSpark 64GB VRam Min): Qwen-Image-2.1-Turbo images and MiniMax H3 / FastH3 video with sound on TensorFold. Apple silicon (Metal) and NVIDIA DGX Spark (CUDA), one install command.
NVFP4 BIZ: nvidia/GLM-5.3-Flash-NVFP4 as distributed, on DGX Spark-class GB10 systems. 2.x serves it with TensorFold (TP=2 or TP=3, FP8 KV, MTP, drafted replies equal serial); 1.x with a pinned vLLM (TP=2 or TP=3, images, optional AXL repack). Apache-2.0 code, MIT weights fetched separately. BIZ = business-use intent, not support or certification.
MiniMax H3 audio-video model on TensorFold NVFP4 kernels: a faster ComfyUI loader for one RTX 50-series GPU (2.3x stock at 20 steps, 81 s per 5 s 1344x768 video with Turbo + sparse attention on an RTX 5070 Ti)
GLM-5.3-Flash EXL3 on 4x RTX PRO 6000 with TensorFold: one-command recipe and engine patches. By Aevonix Research in collaboration with Mia's AI Lab.
GLM-5.3 EXL3 on 4x RTX PRO 6000 with TensorFold: one-command recipe and engine patches. By Aevonix Research in collaboration with Mia's AI Lab.
Kolibri-1 (Aleph Alpha) on two NVIDIA DGX Sparks with TensorFold: ~2x decode tok/s vs one Spark. Recipe, patches, measured results.
GLM-5.3-Flash EXL3 on 2x RTX PRO 6000 with TensorFold: one-command recipe and engine patches. By Aevonix Research in collaboration with Mia's AI Lab.
Keep every AI model on a NAS; insert any of them on your GPU box with one click (Ollama / vLLM) — never evicts a model you didn't name. Stdlib Python, demo mode, web UI.
Unofficial fork of TensorFold v0.6.5 (Python engine line; upstream main is now Zig): EXL3 3.05 bpw + a prompt-lookup (suffix) drafter for Qwen3.8-Flash-Next on CUDA — ~2.4x EXL3 prefill, native vision/video, image + long-prefix caching, every result byte-exact. 14 patches, fixtures, token-hash receipts (provenance: issue 444).
Four DGX Spark switchless TensorFold TP4 deployment for GLM-5.3-Flash EXL3, with ring exchange and measured benchmarks
Qwen3.8-27B AEON Ultimate Uncensored (4-bit MLX) on TensorFold 0.6.0 + DFlash2, Mac Studio M5 Ultra: 191 tok/s single-stream decode, 97 ms TTFT, 1.4k tok/s prefill; one-shot + GHCR carrier
TensorFold v0.6.0 GLM-5.3-Flash engine generalized from 2 to 4 DGX Sparks (TP4) over a switchless RoCE fiber ring - uncensored EXL3 checkpoint, overlay + tooling. Upstreams: ashhart/TensorFold, MiaAI-Lab 2x-Spark stack, alexellis/switchless-nccl.
TensorFold deployment for GLM-5.3-Flash on 4x NVIDIA CMP 170HX (GA100 sm_80): bit-exact 4-rank TP, shm all-gather, W4A16 affine - complete tested solution (199.4 tok/s structured, exactness maintained)
Reproducible MLX Q4 conversion of Cantina Security Apex Flash 1 for TensorFold on dual DGX Spark
Qwen3.8-27B on two NVIDIA DGX Sparks (TP2) with TensorFold: one setup script, benched (99 tok/s single stream, 387 tok/s at 8 users)
GLM-5.3-Flash on TensorFold, 4x RTX PRO 6000: two custom patches, tuned knobs, bench scripts and every measured result (incl. the ones that lost) on top of the Aevonix recipe
Mia's Qwen3.8-Flash-Next recipe for two DGX Sparks (TensorFold Zig engine, NVFP4) plus a CJK draft vocabulary: Chinese decode +38-107%, English -2 to -3%, replies byte-identical.
One command: GLM-5.3-Flash with the Dealign o_proj abliteration transplant on Mia's TensorFold recipe (2x DGX Spark). Byte-verified, stock-speed decode.
GLM-5.3-Flash on 1x RTX 5090 + 2x DGX Spark: attention on the 5090, routed experts on the Sparks over MCDMA RoCE, on TensorFold v0.6.5 (+ 0.6.6's commits) with Mia's GLM work. 2.0 at tag v2.0, v1.0 (glm53f-afd) at tag v1.0. Deployment recipe, as-is.
Helm chart + Ansible to serve GLM-5.3-Flash-EXL3 (EXL3 4bpw, FP8 KV, DFlash2) tensor-parallel on 2× DGX Spark (GB10) joined to an existing k3s cluster via TensorFold v0.6.0
To associate your repository with the tensorfold topic, visit your repo's landing page and select "manage topics."