Sabitlenmiş Tweet
Dan Woods
627 posts
Dan Woods
@danveloper
Vice President of AI Platforms for CVS Health. Former CTO for @JoeBiden.
Katılım Mart 2011
804 Takip Edilen9.5K Takipçiler
@antirez @deepseek_ai I did the same for Qwen3.5-397B - github.com/danveloper/fla… ... one optimization you can do if you want to keep q4 is store them as lz4 - the M3 Max+ has an lz4 hardware accelerator so you can keep more experts hot in the prefill cache. Then expert retrieval is near DRAM speeds.
English
@danveloper Yep consider that this is an M3 Max not an M3 Ultra, in the Ultra I get 2x prefill speed, and the same speed with 4 bit quants instead of 2 bit (only for routed experts, all the other weights are as released by @deepseek_ai).
English
DeepSeek v4 small KV cache + MacBook fast SSD disks = the idea that the disk is not a good target for KV cache is, in this context, totally obsolete. It works *great*. The session you see is opencode using my inference engine for DS4, saving, loading sessions from disk.
English
I'm old enough to remember approx 2 months ago when you could hand Opus 4.6 an arXiv paper and say "let's do this" and it would clunk away for a few hours and deliver a perfect result. Disappointed that's gone, but nonetheless the agentic workflow isn't the problem.
English
My M4 Max MacBook gets 3,756,165 tok/sec in pure C, compared to ~50,000 tok/sec with the FPGA.
Try it yourself: github.com/AlexCheema/tal…
luthira@luthiraabeykoon
We implemented @karpathy 's MicroGPT fully on FPGA fabric. No GPU. No PyTorch. No CPU inference loop. Just a transformer burned into hardware, generating 50,000+ tokens/sec. The model is small, but the idea is not: inference does not have to live only in software 👇
English
@DanielleFong I ran out of tokens until Tuesday (or @thsottiaux saves me), so I bought extra tokens and now I can’t afford fast mode and it’s devastating 😂
English
i use /fast mode with thinking:none / low on chatgpt5.5 and it's awesome.
i call it goblin mode
i wouldn't even think about slow, but gear up to xhigh for big builds.
claude code /fast feels like luxury, or it did. now there's tension between 4.6 /fast and 4.7, and 4.7 is ... great but overconfident. so instead of working on what i want to work on, i am trying ro figure out and work around the model!
i think people who use claude code fast should give goblin mode a try
Jared Friedman@snowmaker
Software engineering job descriptions should really start saying whether they include /fast mode or not.
English
Dan Woods retweetledi
you know what
all of these "which is better" polls are silly
use codex or claude code, whatever works best for you
i am grateful we live in a time with such amazing tools, and grateful there is a choice
English
@danveloper github.com/bigs/deepseek-…
Just wanted to slap it together before I leave my office, but here she is! Haven't read the READMEs yet so don't stab me if it's stupid 😂. But a foundation has been laid! DGX Spark is an annoying target, so it's nice to cobble this together.
English
😩 Guess I'm offline until Tuesday
Dan Woods@danveloper
@thsottiaux y'all got any more of them limits? 🥴
English
@danveloper Well to amuse you while you wait—you inspired me to point codex + /goal at getting deepseek v4 flash running as fast as possible on my DGX Spark. Native FP8/4 hybrid. Just got my first tokens out! Time to optimize :)
English
As far as I know, this always ends well and everyone is happy and safe. We should keep moving forward.
Brian Roemmele@BrianRoemmele
The T-800 is on patrol… Over 13,000 have been ordered.
English