antirez

@antirez • 80,132 subscribers

Reproducible bugs are candies. I like programming too much for not liking automatic programming.

Shorts

32,815 views

18,422 views

16,870 views

14,220 views

Videos

sweetdream.ai

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Private Show

Join now for exclusive access

Free preview available • Premium content

60,229 views • 8 days ago

294,273 views • 1 month ago

319,028 views • 2 months ago

27,006 views • 7 days ago

275,713 views • 3 months ago

63,235 views • 21 days ago

29,067 views • 12 days ago

22,440 views • 11 days ago

149,921 views • 2 months ago

41,545 views • 23 days ago

171,478 views • 3 months ago

26,598 views • 22 days ago

69,587 views • 2 months ago

43,655 views • 1 month ago

26,809 views • 1 month ago

84,127 views • 4 months ago

40,085 views • 3 months ago

44,518 views • 4 months ago

39,023 views • 3 months ago

27,882 views • 3 months ago

Live Cam

antirez

Shorts

Iterating... Evals take time and are boring: but are a fundamental validation step of sane LLM inference. Let's try to make them as easy and fun to run as possible.

DeepSeek v4 Flash has 43 blocks... so, even 4096 batches during the prefill can be smoothly visualized. Look at the new progress bar.

Sicily is the outside of Silo

First steps for a specialized DeepSeek v4 Flash inference engine focused on inference quality / stability at different quantizations, with networked API that is batching capable. This is the 2 bit quants model running on my M3 Max 128GB.

Videos

Watch Anya Live

DeepSeek v4 Flash experimental with vision running (fast) on an m5 max computer and analyzing an image. The implementation for Metal / CUDA / ROCm is finished, just doing the last tests today before releasing it.

That's a bit slow. Streaming directly the K3 official hugging face 1.6TB of weights in mxfp4 in an m5 max 128gb.

GLM 5.2, Q2_K routed experts (effectively ~2.6 bits) running with SSD streaming on an M5 Max 128GB computer.

I'm optimizing tensor parallel execution across two MacBook M5 Max systems using RDMA. This is how fast DeepSeek v4 Flash Exp Vision can run in MXFP4 (full precision) right now *without* DSpark, so this is the min speed you get in practice.

DeepSeek v4 PRO running via SSD streaming on my 128GB MacBook m5 max. 1.6 trillion parameters.

Crazy prefill speed of DwarfStar running DeepSeek v4 PRO in the DGX Station (private branch for now) using techniques that leverage the peculiar RAM + VRAM mixed setup of the Station. 3000 t/s but even more can be reached.

GLM 5.2 flash Q2 (generated by myself with similar recipe of DS4 Flash of DwarfStar GGUF) running on an M5 Max with 128GB. The Q4 weights also work on tensor parallel inference in two M5 Max systems. Soon on GitHub. Still testing + CUDA + ROCm ...

DwarfStar with vision support for GLM 5.3 running on a MacBook m5max computer. The vision support is *not* quantized for max precision for now.

If you need AI to do a search for you in the real world, ds4-agent is basically SOTA, because it can access the web sites without any limitations given that it uses your local Chrome browser (no, not in headless mode, that's the trick...), and DeepSeek v4 is great at search.

DGX Station: this is just the start of what it is possible to do (and why I claim DwarfStar could be *the* Station inference engine). DeepSeek v4 PRO Q2 with routed experts split among VRAM / RAM with kernels optimized for this peculiar setup. 45 t/s but can go faster.

I didn't expect DeepSeek v4 PRO (not Flash) to run well on the Mac Studio M3 Ultra with 512GB of RAM. This is 2 bit quantized with the same DwarfStar recipe used for Flash. 433GB GGUF file. 130 t/s prefill, 13 t/s generation. Prefill in the video is low because small prompt.

DeepSeek v4 PRO running in DwarfStar on a DGX Station. Can go faster, already peaks 50 t/s without DSpark: dynamic VRAM/RAM allocation of experts based on past statistics, 4k t/s prefill on big chunks. VRAM/RAM layer-major streaming. Mixed RAM/VRAM inference requires care.

First kinda working implementation of GLM 5.2 in DwarfStar. Will take some time to be good enough, but it is a promising start. 433 GB GGUF file. M3 Ultra 512GB.

Still didn't attempt any quality testing, but let's say that GPT 5.6 Sol was able almost unassisted to write this Laguna S2.1 inference implementation following the other two models in DwarfStar, and it is very fast, 50 t/s generation, very fast prefill as well.

Only 10 steps in this example to make the video short, yet H3 is quite fun.

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

For the DGX Spark owners. This is what you get with DS4 in your hardware. I want to post this to show how with fast prefill and not very fast generation, the system remains absolutely fine to use.

In case you have doubts about the q2 quants inference of DS4 (I noticed many don't trust the README claims), here is it analyzing Picol source code using the pi agent.

You didn't see this coming in ds4-agent, did you?

The first M5 max arrived! Many many thanks to our sponsors ⿻ Audrey Tang 唐鳳 and Niels Gron, the next will arrive on Monday.

I'm optimizing tensor parallel execution across two MacBook M5 Max systems using RDMA. This is how fast DeepSeek v4 Flash Exp Vision can run in MXFP4 (full precision) right now without DSpark, so this is the min speed you get in practice.

DwarfStar with vision support for GLM 5.3 running on a MacBook m5max computer. The vision support is not quantized for max precision for now.

DGX Station: this is just the start of what it is possible to do (and why I claim DwarfStar could be the Station inference engine). DeepSeek v4 PRO Q2 with routed experts split among VRAM / RAM with kernels optimized for this peculiar setup. 45 t/s but can go faster.