YASAKANI is built for Windows. I got it running on Linux (Ubuntu under WSL2) with one compiler flag.
RTX 5090, Qwen3.8 27B:
1,277 tok/s on a code edit, same answer as plain llama.cpp at 76
597 tok/s on new code
Its self-check passes.
Windows or Linux for local AI?
I ran it on my 5090: Both are compressions of the same Qwen3.8-27B: your quality file (12.4 GB) vs NVIDIA's NVFP4 (19 GB: 4-bit MLPs, 8-bit attention) on vLLM using the same prompts + greedy.
Tool calling (350 BFCL questions): as served, @XaeaZii is right, 78% vs 86%. But it's a parser bug, not the weights. The model sometimes writes a boolean Python-style (True), and the server passes it on as the text "True". With that fixed it's 87% vs 86%. Fix and numbers in a PR: https://t.co/VVHWHLjTzB
The rest came out level:
- GSM8K math (500): 96.4% vs 96.4%
- BFCL multi-turn (60 conversations): 67% (73% with the fix) vs 70%
- 256K context: both found all 9 hidden passphrases; blind-graded retellings 22/27 vs 19โ22/27
On the 1,200 tok/s at 256K: NVFP4 fits the full 262K window on one 5090 (8-bit KV) but writes ~53 tok/s, ~100 with its MTP drafter. Nothing I ran hit 1,200 at 250K.
YASAKANI is built for Windows. I got it running on Linux (Ubuntu under WSL2) with one compiler flag.
RTX 5090, Qwen3.8 27B:
1,277 tok/s on a code edit, same answer as plain llama.cpp at 76
597 tok/s on new code
Its self-check passes.
Windows or Linux for local AI?
That's a fair question, and speed is pointless if the answers get worse, so I checked that.
The speed doesn't come from cutting corners on the answer. A small guesser drafts the next few words and the full model checks every one before keeping it, so with the guesser on or off you get the same text. On the code edit in my post, plain llama.cpp at 76 tok/s and YASAKANI at 1,277 wrote the identical answer.
On quality, I ran small test sets on YASAKANI's model and the standard version of the same model: 20 coding problems, 20 maths problems and 200 general-knowledge questions. They scored 19, 18 and 160 against 19, 18 and 163. That's no measurable loss on sets that size.
Where speed really pays is thinking models: they write thousands of words of reasoning before the answer, so a 4x faster engine is the difference between letting it think and turning thinking off.
It's a 27B open model on one home PC, so it won't match the frontier models. But it's private, free to run, and now quick enough to use.
@yasakani_dev Thank you for merging it so quickly, and for the credit! I've sent the Linux build steps for BUILDING.md as PR 10 (tested on WSL2 / Ubuntu 24.04 with my RTX 5090).
Thanks for releasing YASAKANI!
It builds and runs on Linux (Ubuntu 24.04 under WSL2) with one flag: -D_fseeki64=fseeko -D_ftelli64=ftello. Selfcheck passes on all 3 prompts. Fastest model, my 5090: 1,277 tok/s on a code edit, 597 on new code. Script: https://t.co/YAxVtfSuwo
Linux๏ผWSL2 ใฎ Ubuntu๏ผใงใๅใใพใใใๅ ฌ้ใใใใจใใใใใพใ๏ผ
Full test, 13 min: https://t.co/B1zVB7mEim
Linux build script + raw results: https://t.co/L4fKm5U8H8
YASAKANI by @yasakani_dev: https://t.co/rqmleZtjiG
Did NVIDIA lock your RTX 3090 at half speed?
A free open-source tool lifts a GeForce limit on RTX 30 cards: 82 -> 155 TFLOPS in one owner's test. Real AI work: about 10-60% faster.
Experimental, see video below:
Wake up.
Every profession deemed white-collar, the middle class dream, is about to be decimated by AI, at a much faster pace than anticipated, with monstrous social consequences.
as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better. security leaders, never let a crisis go to waste: please go to your boards and ask for help right now. not for ai, for the basics. (we need to hold on for ~18 months.)
AI just learned COBOL, the 1959 language still behind a lot of your bank's code.
On OpenAI's own test, its latest model's COBOL rewrites passed every hidden test, byte for byte, 86.6% of the time (97.9% with more thinking).
Should banks panic? Not yet. Here's the catch ๐
@SashaNaryshkine I think that's exactly the problem because every system has its own rules and they are often undocumented. A code rewrite has to match it exactly. Somewhere someone is quietly getting an extra penny!
I turned the new Roadster ramp-flip stunt into a pixel game ๐
How many flips can you land? ๐
Built with @AnthropicAI Opus 5.5, inspired by @Tesla, hosted on @Cloudflare