Two 300B-parameter models. One 128 GB mini PC.

Yamz Labs quantizes large open models and builds the engine to run them on a single AMD Ryzen AI Max+ 395 machine with 128 GB of memory. No server, no cloud: the weights, the engine and the benchmarks are public.

Models on Hugging Face Kyojin on GitHub

First two models

GLM-5.3-Flash

320B parameters, 18B active. EXL3, built from turboderp's public quantization with our layer mix and tuning.

99.7 GBSize

580 tok/sPrefill

26 to 30 tok/sDecode

Top-1 agreement with the official FP8 model: 90.11 % on 126 held-out rows. Prefill measured at 3.5K, 14K and 64K tokens.

Model card

MiMo-V2.6-Flash-MOPD

309B parameters, 15B active. EXL3 quantization.

105 GBSize

~650 tok/sPrefill

up to 44 tok/sDecode

Top-1 agreement with the official FP8 model: 91.99 %. Decode uses speculative decoding: 32.1 on prose, 34.8 on chat, 44.3 on code.

Model card

Prefill and decode speed of GLM-5.3-Flash and MiMo-V2.6-Flash on one Strix Halo machine

KLD and top-1 agreement against the official FP8 model

All figures: AMD Ryzen AI Max+ 395, 128 GB, one machine. Quality is measured against the official FP8 weights. Uncensored variants of both models are also published. Methods and limits are on each model card.

Kyojin engine

Kyojin runs these models on AMD Strix Halo (gfx1151) with ROCm. It is built on ExLlamaV3 by turboderp, whose EXL3 format and quantization work make this possible. Our thanks to turboderp and the ExLlamaV3 contributors. Kyojin adds the kernels, memory handling and speculative decoding that these models need on unified-memory hardware.

github.com/Yamz-Labs/kyojin

For companies

The same work, applied to a business: a private model installed on your own machines, automation of repetitive office tasks, and websites. Your data stays inside the company. Services are delivered in France by GETA Solutions Informatique, with entry prices shown.

getasolutions.fr/informatique (in French)

First release

More models and speed work are coming. Contributions and benchmarks are welcome: run a model on your own machine and tell us your tokens per second and hardware.