How I run the models mentioned here locally?
- GLM 5.3 Flash --> 2x RTX PRO 6000s
- Qwen3.8-Flash-Next --> 2x DGX Sparks
- DeepSeek V4.1 Flash --> 4x DGX Sparks
GLM 5.3 Flash is the main worker.
Big update for GLM 5.3 Flash for 2x DGX Sparks
Running latest TensorFold (v1.3.0). Lots of changes:
- Faster decode: CUDA kernels now chain directly with fewer launch gaps (PDL)
- Cleaner stack: Our paged FP8 KV, MoE, and DFlash2 speedups are now built into upstream (100+ patches down to 13)
- Smarter prompt cache: Cache space is reserved up front, and multi-turn thinking chats reuse cached reasoning instead of re-reading history
- More stable 2-GPU vision: Long image prompts no longer time out the second GPU
Even more changes and improvements coming soon!
Get it here:
https://t.co/Xein7tYXH3
Running two DGX Sparks on one cable? Check NCCL_IB_HCA.
Each Spark's port shows up as two network devices, and one alone gets about 112 of the 200 Gb/s.
List both and long prompts process 26-28% faster. Generation speed stays the same.
https://t.co/4IZx8dXVBD
WorkBuddy AI Class · EP01:
Drop a long meeting recording into WorkBuddy. Get the transcript, meeting minutes, follow-ups, and a client profile—all in one go.
See it in action below 👇
Run @TencentHunyuan 220B Hy4 Flash locally on a 64 GB machine
The trick: 63.5B of the model is an Engram lookup table read a few KB per token, so it stays on your SSD
Get quants here 👇
https://t.co/LuAMG77hSp
This is the type of things you're see a lot more going forward. Improving existing open-weight models.
RadixArt released DaisyRose-27B 🌹, a trained Qwen3.8-27B focused on long horizon agentic tasks.
A new beginning for local AI on Windows 💻
Run open models with Atomic Chat on NPU, GPU & CPU using Foundry Local. Keep agents sandboxed with @Microsoft Execution Containers (MXC)
Private. Local. Sandboxed.
Public Beta: Early Dec 2026
GA: Jan 2027
Get access: https://t.co/15gfisHOTy
hot take: `sudo reboot` is a performance optimization
a stopped Spark: ~25 GB free in blocks the GPU can actually use. after dropping caches and compacting: ~80 GB.
`free -h` shows you none of this. memory fragmentation is real, it's hidden, and it's a problem.
reboot before you start your inference engine. "turn it off and on again" was right the whole time.
https://t.co/A77uVCpQYm
if you have a dgx spark, give qwen3.8 27b at 6.0 bpw a try.
it's the closest thing to the full bf16 model that still fits on one spark. the one to pick when quality matters more than speed.
try it and tell me how it runs on your box. link in the first comment.