What Xolotl is
Xolotl runs DeepSeek-V4.1-Flash – all 552 billion parameters – on one
gaming PC at 33 tok/s on real agent tasks (41 tok/s in fast mode), next
to GLM-5.3-Flash, Qwen3.8-Flash-Next, Qwen3.8-27B, LFM3-Omni and Muse
Glimmer 30B, all behind a single OpenAI- and Anthropic-compatible
server. Linux, NVIDIA, MIT.
Built on Strata, then grown
Niko1221’s engine, server and core kernels are still at its heart
(merged at 0.1.44.1), but Xolotl adds more than 70,000 lines on top: its
own DeepSeek-V4.1, V4-Flash and GLM-5.3-Flash engines, a Qwen3.8-27B
dense engine grown from Mirko Perrone’s port, the LFM3-Omni engine, fast
mode, Xolotl Direct, and the quantisation, expert-tiering and serving
work behind every model here.
Numbers, measured – and usually default on
The changelog is specific and reproducible. Qwen3.8-27B with a DFlash2
drafter moved from 171 to 204 tok/s on a best-case 2K prompt (a 13.7 GiB
XQ quant, one RTX 4090). Muse Glimmer 30B runs natively at its full 512K
context on one 24 GB card at 47 tok/s decode, up to 379 tok/s with
Meta’s drafter and 3,100 tok/s prefill. Many of those gains arrive
labelled "same text, bit for bit; default on" – a performance change
that leaves the output identical. Some builds even read row tables
measured on the exact card you run, not a nearby model.
Why it matters to you
It is the practical end of the Mixture-of-Experts idea: a frontier model
on hardware you can buy at a games shop, served through the same API
your tools already speak.
We install and benchmark this class of build on your own rig across
Perpinyà, Girona and Barcelona. Ask: privatelocalai@proton.me. Source:
https://github.com/Indras-Mirror/Xolotl
Top comments (0)