Update for GLM 5.3 Flash for 2x DGX Sparks ✨ - Tool calling improvements: Tool eval bench hard mode score is now 93 with thinking on! - Long agent sessions stay cached: a 380k-token chat resumes in 3s instead of 4min (!) - Video input now work with WITH_VIDEO=1. - Fixed a JSON bug where an ID came back with a digit missing. Highly recommend you to update ASAP! Get it here:

Local AI will set you free. No usage limits. No price hikes. No model getting quietly worse overnight. You pick the model, run it & it stays the same. Your prompts never leave your machine. Own your intelligence. Fully private.

Run the NEW GLM 5.3 Flash recipe (1104) on your 2x NVIDIA DGX Sparks with ease ✨ Based on my RTX PRO 6000 recipe and ported to GB10. 93 of its patches are unchanged, and 12 are unique GB10 patches. Your sessions will run smoother, feel snappier, and in general be a more pleasant experience from top to bottom. Focused on uncompromising quality. Everything has been tested & verified, and the default settings are the best quality you can run. Native 1M context, 8 streams, TensorFold. Decode performance (non-greedy prose): ~ 48 tok/s single stream. ~ 87 tok/s at 4 concurrent streams. - You will feel a noticeable difference vs the older recipe. Prefill: ~1,700-1,800 tok/s, 4k-256k. I ran all benchmarks with the GPU clock cap set to 2200. Support for 4x DGX Sparks is coming soon! The link to the new recipe is below.

Opus 5.5 has been downgraded to B. It's still awesome in frontend, but for me it's almost unusable for the type of backend work I do. Too many mistakes. S+ is empty, waiting for a new king.

DeepSeek V4.1 Flash 1017 for 4x DGX Sparks got a lot of updates, with support for @NVIDIAAI's NVFP4, and a lot PRs/issues merged/fixed. NVIDIA's NVFP4 - Can be enabled by setting NVFP4=1. - 4-6% increase on prefill speeds. - A bit faster decode on multi-concurrency. ✨Faster loading - Each worker can load the model from its own disk instead of over the network (opt-in); reported load time from 315s to 96s. - The weight file is read ahead of the loader; reported node load time from 83s to 27s. ✨More reliable replies - A reply format rule that can't start now fails only its own request. - A failed first GPU launch leaves the cached graph safely in its cache. - JSON answers stay JSON instead of being mistaken for tool calls. - A reply containing only stray tool-call closing tags now gets a warning, not a guessed repair. ✨Safer scripts - Long-running background processes no longer hold on to the launcher's locks. - On switched networks, nodes prefer their shared fast network when connecting at startup. - ./stop.sh can now warn about requests still in progress, and the ./start.sh test needs a real model reply. - IPv6 addresses now work in the URLs ./start.sh checks and prints. - Preparation tools now run on the same image as the server. - ./prepare.sh now stops if indexing fails, instead of deleting the weight file. Updates also coming for 2x DGX Spark version! Thanks to shalk3, @SvangenStudios, @tfolkman, Christopher A. Williamson (cawilliamson), and Toshiaki Maki. Link below.

BREAKING: GLM 5.5 training traces show they might ship 32B, 122B and 256B variants 🔥🔥 If they DO ship a 32B, Qwen3.6-35B might finally be beaten! The next two weeks should be interesting 👀

Absolutely insane decode numbers on 2x RTX PRO 6000s with GLM 5.3 Flash on TensorFold. Haven't even optimized it really... faster than API I think. This is with power capped at 400w per card.

It's 100% worth capping the DGX Sparks. Almost no performance hit, but WAY cooler temps. All of my sparks are capped at 2200. You can use sparkLimiter or sparkDash to easily do it.

Wild idea. The UK does not have any AI labs creating models. Who would be up for banding together in the UK to see if we can create a open community model 🤷‍♂️ Maybe it will be trash but I for one would love to learn and try. Dm me if you are game 💪🏻

GPUs and memory prices are at ATH, and I don't see any reason for them to slow down anytime soon. Owning intelligence has become a high luxury. Those who call AI a bubble really have no idea. The "bubble" will come at some point, but we're nowhere near there. And all those anti AI people who are "sick of it" are going to be left behind and be sorry later that they missed the train. You either join it, or others will beat you. SIMPLE. Evolution at its best.

I'd take 4x DGX Sparks over M5U 512 GB any day! The system is so mature right now, and works so well, it will take YEARS for the Mac to get this good, if ever. You can't get with one Mac the versatility you have with 4 Sparks. You just can't. I can decide to run 2 models per 2 Sparks, or 1 model on all 4 Sparks, or 1 model on one Spark and another model on 3 Sparks. What about diffusion models? Image and video generation. How will that work if you run image gen on your Mac but also want to use an LLM? The versatility of the DGX Sparks is insane! Also, CUDA.