Project Maya v1.1.16 🐳 - DeepSeek-V4.1-Flash-1017 (748B) on your own PC - Maya-S, our own quant: 178 GB, 92.6% same next token as the original - NVIDIA + AMD, Linux + Windows - Experts from RAM/SSD: runs below its size

3090 has never seen this before🔥 Got DeepSeek-v4.1-flash-1017 running on my NVIDIA RTX 3090 + 64GB RAM + SSD Using Maya-S quant by @PeasantSmith: 178.2 GB of Maya-S + 202.8 GB of Engram = ~381 GB FP4: KV Project Maya inference engine Speed is slow btw, but it's shocking and fun to see a model as big as that running on my machine. Avg decode: 4t/s prefill - 8k: 121 tok/s - 16k: 183 tok/s - 32k: 230 tok/s - 64k: 243 tok/s I'll keep a track of progress as we get more updates, so expect more posts soon.

350M tokens!!!! And the quality is fantastic 🤩 GLM 5.3 Flash really is a thing

Just launched Turret Control: Operation Lumberjack — a tank-on-rails wave shooter set in the 1945 European Theatre. Play it here: arcade.digitalspaceport.com Built in ~1 week with Project Maya + GLM 5.3 Maya-L (1M context, ~350M tokens @ 55 tok/s gen / 1250 tok/s prompt)

@PeasantSmith Project Maya v1.1.17 posts a number my cards can use: GLM-5.3-Flash long prompts got faster on Volta. A 27K prompt went from 41.3 to 33.4 seconds once the routed prompt experts became FP16 cuBLAS GEMMs and the DSA indexer moved onto the tensor cores. Both my V100s are Volta, and GLM at 262K is what I run on them. The loud number is DeepSeek's: 8K prompt chunks instead of 16-token windows took a 32K prompt on 7x3090 from 99 to 1,010 tok/s! Currently testing Minimax M3.1-Flash, but DeepSeek on Maya is in the testing queue!

Project Maya v1.1.17 - DeepSeek-V4.1-Flash: prompts in 8K chunks, 1000+ tok/s prefill on multi 3090 - Maya-S: one 340 GB set from our repo, Engram included - GLM-5.3-Flash: repetition penalties, faster long prompts on Volta - 17 community PRs merged github.com/mw00/project-maya…

One of the first gens with local Deepseek V4.1 Flash 1017 in Maya So far so good. 4x3090 - 256k ctx Around 35-45 t/s with prefill 500-1000t/s More testing tomorrow. @PeasantSmith - good job :) ps. it's Almost 400GB model - with kisses to GB10 friends :P

Holy prefill!!!!!! And most definitely we can get that speed up very soon 🫡⚡

Deepseek-v4.1-Flash is now available in Project Maya - and it’s fast right out of the gate! Testing at 256k context: 🔥Prefill: 2,000 tps 🚀Decode: 50 tps A little bit slower compared to GLM-5.3-Flash on the same hardware, I bet there’s still optimization headroom.

Deepseek 1017 this is the same prompt I used with Opus and a few other models.. Deepseek is definitely the best Open model result by far, rivaling Opus result.

Two frontier open weight models - GLM and DeepSeekFlash - can now run locally on my own hardware, faster with each and every engine release. No API. No bill. Intelligence on tap to power my agents. Check out Project Maya - and follow @PeasantSmith

Don't forget to leave that follow! Many more updates are already full steam and you don't want to miss it 🫡

DeepSeek4.1-Flash n-gram memory is a giant learned lookup table keyed by short token sequences, and it's the reason this model drags 203 GB of tables around. It's why the model is cheap to run but big on disk: the lookup costs 24 SSD reads per token instead of billions of FLOPs.

@superalesha lets fucking go!

Just in time 🫡 run DeepSeek today on your own hardware with Maya 👇 github.com/mw00/project-maya

Monday: WE NEED GPUs! Tuesday: WE NEED RAM! Wednesday: WE NEED STORAGE! What will be next?😅😂🤣 Downloading Deepseek 4.1 Flash... just around 400GB.

Back to work on PRs and issues now