Repository navigation
Releases
XeMaya 0.1.0-beta.1 (Linux, Intel Arc)
Native inference for Intel Arc GPUs on Linux (SYCL / Level Zero, XMX), for Qwen3.8-Flash-Next, GLM-5.3-Flash and
DeepSeek-V4.1-Flash. XeMaya is a fork of XeStrata by MistVVK, itself a port of
Strata by Niko1221, and targets the models and quantisations of
Project Maya by mw00.
The first beta. The alphas asked whether it runs; since then each of the three model families has been run again
on GPUs against a reference or a recorded benchmark after the changes below. It is still for Linux and Intel Arc
only, and what was not checked is listed under Known limitations.
What changed since 0.1.0-alpha.2
-
DeepSeek-V4.1-Flash reads a prompt much faster when the model is larger than the VRAM. The regular 384-expert
IQ2_XXS file (123 GiB of experts) on an Arc Pro B70 + B65:Prompt 0.1.0-alpha.2's code This release's code The 63 tokens after it 576 tokens 20.0 tok/s 36.3 tok/s 12.8 -> 14.7 tok/s 2,048 tokens 36.5 71.6 11.3 -> 16.2 8,192 tokens 35.2 91.8 8.4 -> 15.9 The generated tokens are the same. Three changes: an expert the VRAM does not hold is read from the files
straight into pinned memory, the VRAM tier's fill after a prompt reads 16 experts side by side, and the experts a
prompt's chunk uploads stay in the tier's free slots (an 8,192-token prompt read 804 GB from the drive before,
218 GB now). Each has a switch:XEMAYA_PROMPT_PINNED_READS=0,XEMAYA_TIER_FILL_BATCH=0,XEMAYA_PROMPT_KEEP=0. -
Qwen3.8-Flash-Next decodes faster with drafts: 77.3 -> 79.4 tok/s on code and 58.0 -> 59.9 on prose after a
26.8K-token prompt (B70 + B65); Strata v0.2.12 measured 78.7 / 59.2 in the same comparison. The draft layer's
attention runs on the matrix engines and the draft head's 2-bit rows in bit planes. -
Opt-in:
XEMAYA_EXPERT_UNBUFFERED=1reads a prompt's experts past the file cache (32.8 / 60.3 / 60.8 tok/s at
the three lengths before the kept experts; the first tokens after a short prompt come slower). -
GLM-5.3-Flash prompts whose experts come from the files take the same faster path (a 198-token prompt: 20.3 s ->
12.2 s on B70 + B65).
Install
sha256sum -c SHA256SUMS --ignore-missing
tar -xf xemaya-0.1.0-beta.1-linux-x86_64.tar.xz && cd xemaya-0.1.0-beta.1
./xemaya.sh setup
./xemaya.sh doctor
./xemaya.sh models /path/to/your/models
./xemaya.sh serve <id> --port 8095 --no-api-keyNeeds an x86-64 CPU with AVX2, an Intel Arc GPU with Intel's compute runtime (Level Zero), glibc 2.38 or later
(Ubuntu 24.04+, Fedora 40+) and Python 3.10 or later. setup installs Intel's SYCL and oneMKL run-time libraries
from PyPI into the folder. Details: INSTALL.md.
Recorded results
Arc Pro B70 (32 GB) + B65 (32 GB), Intel oneAPI 2026.2, development builds of this release's code on 2026-11-11 and
2026-11-12 unless the row says otherwise. Each row is one file, one prompt set and one configuration; they do not
compare with each other.
| Model and file | Prefill | Decode | Agreement with the llama.cpp CPU reference |
|---|---|---|---|
| DeepSeek-V4.1-Flash regular (384 experts) GSQ-RCO IQ2_XXS, B70 + B65 | 36.3 tok/s at 576 tokens, 91.8 at 8,192 | 14.7–16.2 tok/s | 16 of 16 forced tokens on the 576-token reference |
| DeepSeek-V4.1-Flash Coder GSQ-RCO IQ2_XXS, B70 + B65 | 52.1 tok/s at 576 tokens (the first prompt of a process) | 16.7 tok/s | 16 of 16 forced tokens on the 576-token reference |
| GLM-5.3-Flash Maya-S v2 IQ2_XXS, B70 + B65, 198-token prompt | 16.3 tok/s (the first prompt of a process) | 12.8 tok/s | 125 of 128 on the chat-agent reference |
| Qwen3.8-Flash-Next GSQ-RCO IQ3_S, B70 + B65, 26.8K-token prompt | 1,864–1,940 tok/s | 79.2–79.4 / 59.9 tok/s (code / prose) | Speed only; Strata v0.2.12 on the same cards: 78.7 / 59.2 decode |
Earlier three-GPU results (DeepSeek Coder 207–209 tok/s prefill, GLM 152–161 at 2,377 tokens) are in the
README; they were not run again.
How this release was checked
- Built from a clean export of the tagged commit with
tools/xemaya/release.sh; its host tests and the GLM, DeepSeek
and MiniMax kernel parity tests on the OpenCL CPU device pass (MiniMax's is not shipped). - Installed from the archive in a fresh Ubuntu 24.04 container: checksums,
setup, the libraries resolve, the
CPU-device parity tests pass, the server answers against its mock engine. - On GPUs: the published build's own check (DeepSeek-V4.1-Flash, its Coder file and GLM-5.3 against their
references on B70 + B65) is recorded at the end of these notes on the release page. Qwen3.8 was not run with
the release build: its numbers above are from a development build of the same code.
Known limitations
- Reference agreement is recorded for the files and prompts above only. Other quantisations and long contexts are
not qualified. - GLM-5.3-Flash Maya-S v2 agrees with its chat-agent reference at 125 of 128 forced tokens, one fewer than in the
alphas, with the same median and maximum difference: the VRAM tier holds other experts after the prompt, and near
ties fall the other way.XEMAYA_PROMPT_KEEP=0gives the earlier tier. - A model larger than the VRAM still reads the experts that do not fit once for every 1,024 tokens of a prompt.
- The release is built with Intel oneAPI's icpx. Builds with intel/llvm, the three-GPU configurations and the
slow-link card were not checked again for this release. - MiniMax-M3.1-Flash is not in this release: its code is in the source and is held back (slow, and one development
run ended in a GPU engine reset of unknown cause). - DeepSeek-V4.1's tool calls are read by the server but were not exercised with the model; its thinking is not
qualified. - A GPU on a slow or shared link can be starved by long expert streaming.
- Experimental and unsupported in this release: the hybrid segment (a model's last blocks on another machine), the
Windows build with the CUDA target (compiles, never run), NVIDIA and AMD GPUs, the inherited deb/rpm packaging and
setup.pyinstaller.
Files
| File | |
|---|---|
xemaya-0.1.0-beta.1-linux-x86_64.tar.xz |
The engine built for Intel GPUs, the server, the xemaya.sh command, and the complete source of the build |
xemaya-0.1.0-beta.1-source.tar.gz |
The source alone |
SHA256SUMS |
Checksums of both |
Licence: LGPL-3.0-or-later, with MIT code from Strata, Project Maya and ggml / llama.cpp
(NOTICE).
Maintainer: Jumbomuffin777.
The published build on GPUs (2026-11-12)
Installed from this archive with the PyPI run-time on Arc Pro B70 + B65, each against its llama.cpp reference:
DeepSeek-V4.1-Flash regular (384 experts) IQ2_XXS 16 of 16 forced tokens (median |d logprob| 0.012, max 0.097; 576-token
prompt read at 34.1 tok/s), its Coder file 16 of 16 (0.010, 0.155; 56.1 tok/s), GLM-5.3-Flash Maya-S v2 125 of 128 on
chat-agent (0.012, 0.395). No kernel message.
9b48f6c28df1ce5e02107ae761b6416a011918f2eab215b5ce9317ae5a33e78b xemaya-0.1.0-beta.1-linux-x86_64.tar.xz
f74c26b7a242aa5a64b64efdcdbd1556311f952a67513e7f1e79a716ac06dfa5 xemaya-0.1.0-beta.1-source.tar.gz
XeMaya 0.1.0-alpha.2 (Linux, Intel Arc)
XeMaya 0.1.0-alpha.2
The first public release of XeMaya: native inference for Intel Arc GPUs on Linux (SYCL / Level Zero, XMX),
for Qwen3.8-Flash-Next, GLM-5.3-Flash and DeepSeek-V4.1-Flash.
XeMaya is a fork of XeStrata by MistVVK, itself a port of
Strata by Niko1221, and targets the models and quantisations of
Project Maya by mw00.
What changed since 0.1.0-alpha.1
- Fixed: the server no longer fails to start for a DeepSeek-V4.1 model whose pack's chat template is not valid
UTF-8; it falls back to the V4.1 template and says so. Thanks to @bettercallcaleb for the fix
(#1), XeMaya's first outside contribution. - Recorded speeds of the regular DeepSeek-V4.1-Flash file (384 experts) were added to the repository.
This is an alpha. It is for Arc owners who want to try it and report what happens, not for anything you depend on.
Install
sha256sum -c SHA256SUMS --ignore-missing
tar -xf xemaya-0.1.0-alpha.2-linux-x86_64.tar.xz && cd xemaya-0.1.0-alpha.2
./xemaya.sh setup
./xemaya.sh doctor
./xemaya.sh models /path/to/your/models
./xemaya.sh serve <id> --port 8095 --no-api-keyNeeds an x86-64 CPU with AVX2, an Intel Arc GPU with Intel's compute runtime (Level Zero), glibc 2.38 or later
(Ubuntu 24.04+, Fedora 40+) and Python 3.10 or later. setup installs Intel's SYCL and oneMKL run-time libraries
from PyPI into the folder. Details: INSTALL.md.
Recorded results
Arc Pro B70 (32 GB) + B65 (32 GB) + B60 (24 GB), Intel oneAPI 2026.2, 2026-11-09 and 2026-11-10, development builds
of earlier commits. Each row is one file, one prompt set and one configuration; they do not compare with each other.
| Model and file | Prefill | Decode | Agreement with the llama.cpp CPU reference |
|---|---|---|---|
| DeepSeek-V4.1-Flash Coder GSQ-RCO IQ2_XXS, three GPUs | 207–209 tok/s at 576 tokens, 225 at 2,876 | 19.5–21.7 tok/s | 16 of 16 forced tokens on both references |
| GLM-5.3-Flash Maya-S v2 IQ2_XXS, three GPUs, 2,377-token prompt | 152–161 tok/s | 17.4 tok/s (20.0 with MTP drafts) | 126 of 128 on the chat-agent reference |
| Qwen3.8-Flash-Next GSQ-RCO IQ3_S, B70 + B65, 26.8K-token prompt | 1,842–1,884 tok/s | 76.6 / 57.1 tok/s (code / prose) | Speed only; Strata v0.2.12 on the same cards: 2,211 prefill, 78.3 / 59.2 decode |
On the same GLM file and prompt, llama.cpp's SYCL backend measured 68.3 tok/s prefill and 13.6 tok/s decode on this
machine. Qwen is behind upstream Strata, as the table says.
How this release was checked
- Built from a clean export of the tagged commit with
tools/xemaya/release.sh; 13 host tests and the GLM, DeepSeek
and MiniMax kernel parity tests on the OpenCL CPU device pass (MiniMax's is not shipped). - The fix:
serve/test_dsml.py, 26 passed; its new test fails without the change. CI (the server's tests, licence
lint) passes on the tagged commit. - Installed from the archive in a fresh Ubuntu 24.04 container: checksums,
setup, the libraries resolve, the
CPU-device parity tests pass, the server answers against its mock engine. (0.1.0-alpha.1, the same engine source,
was also installed in Ubuntu 26.04 and Fedora 44 containers.) - On GPUs: the published build's own check is recorded at the end of these notes on the release page. GLM-5.3 and
Qwen3.8 were not run with the release build: their numbers above are from development builds.
Known limitations
- Reference agreement is recorded for the files and prompts above only. Other quantisations and long contexts are
not qualified. - MiniMax-M3.1-Flash is not in this release: its code is in the source and is held back (slow, and one development
run ended in a GPU engine reset of unknown cause). - DeepSeek-V4.1's tool calls are read by the server but were not exercised with the model; its thinking is not
qualified. - Qwen3.8 and GLM-5.3 were not rechecked on GPUs after the last shared expert-tier changes.
- A GPU on a slow or shared link can be starved by long expert streaming.
- Experimental and unsupported in this release: the hybrid segment (a model's last blocks on another machine), the
Windows build with the CUDA target (compiles, never run), NVIDIA and AMD GPUs, the inherited deb/rpm packaging and
setup.pyinstaller.
Files
| File | |
|---|---|
xemaya-0.1.0-alpha.2-linux-x86_64.tar.xz |
The engine built for Intel GPUs, the server, the xemaya.sh command, and the complete source of the build |
xemaya-0.1.0-alpha.2-source.tar.gz |
The source alone |
SHA256SUMS |
Checksums of both |
Licence: LGPL-3.0-or-later, with MIT code from Strata, Project Maya and ggml / llama.cpp
(NOTICE).
Maintainer: Jumbomuffin777.
The published build on GPUs (2026-11-11)
Installed from this archive with the PyPI run-time on Arc Pro B70 + B65 + B60: DeepSeek-V4.1-Flash Coder IQ2_XXS matched the llama.cpp reference at 16 of 16 forced tokens on both saved prompts (median |d logprob| 0.009 and 0.010), no kernel message. CI on the tagged commit: passing.
bad99fb21973a7f204c7b207e462b188177a4dd92de19fdc62a30b3f8666f288 xemaya-0.1.0-alpha.2-linux-x86_64.tar.xz
1564aac23afda343630bd5da068b0afb3ad57b729680ab2bdc148d73f64a1b9e xemaya-0.1.0-alpha.2-source.tar.gz
Contributors
Assets 5
XeMaya 0.1.0-alpha.1 (Linux, Intel Arc)
XeMaya 0.1.0-alpha.1
The first public release of XeMaya: native inference for Intel Arc GPUs on Linux (SYCL / Level Zero, XMX),
for Qwen3.8-Flash-Next, GLM-5.3-Flash and DeepSeek-V4.1-Flash.
XeMaya is a fork of XeStrata by MistVVK, itself a port of
Strata by Niko1221, and targets the models and quantisations of
Project Maya by mw00.
This is an alpha. It is for Arc owners who want to try it and report what happens, not for anything you depend on.
Install
sha256sum -c SHA256SUMS --ignore-missing
tar -xf xemaya-0.1.0-alpha.1-linux-x86_64.tar.xz && cd xemaya-0.1.0-alpha.1
./xemaya.sh setup
./xemaya.sh doctor
./xemaya.sh models /path/to/your/models
./xemaya.sh serve <id> --port 8095 --no-api-keyNeeds an x86-64 CPU with AVX2, an Intel Arc GPU with Intel's compute runtime (Level Zero), glibc 2.38 or later
(Ubuntu 24.04+, Fedora 40+) and Python 3.10 or later. setup installs Intel's SYCL and oneMKL run-time libraries
from PyPI into the folder. Details: INSTALL.md.
Recorded results
Arc Pro B70 (32 GB) + B65 (32 GB) + B60 (24 GB), Intel oneAPI 2026.2, 2026-11-09 and 2026-11-10, development builds
of earlier commits. Each row is one file, one prompt set and one configuration; they do not compare with each other.
| Model and file | Prefill | Decode | Agreement with the llama.cpp CPU reference |
|---|---|---|---|
| DeepSeek-V4.1-Flash Coder GSQ-RCO IQ2_XXS, three GPUs | 207–209 tok/s at 576 tokens, 225 at 2,876 | 19.5–21.7 tok/s | 16 of 16 forced tokens on both references |
| GLM-5.3-Flash Maya-S v2 IQ2_XXS, three GPUs, 2,377-token prompt | 152–161 tok/s | 17.4 tok/s (20.0 with MTP drafts) | 126 of 128 on the chat-agent reference |
| Qwen3.8-Flash-Next GSQ-RCO IQ3_S, B70 + B65, 26.8K-token prompt | 1,842–1,884 tok/s | 76.6 / 57.1 tok/s (code / prose) | Speed only; Strata v0.2.12 on the same cards: 2,211 prefill, 78.3 / 59.2 decode |
On the same GLM file and prompt, llama.cpp's SYCL backend measured 68.3 tok/s prefill and 13.6 tok/s decode on this
machine. Qwen is behind upstream Strata, as the table says.
How this release was checked
- Built from a clean export of the tagged commit with
tools/xemaya/release.sh; 13 host tests and the GLM, DeepSeek
and MiniMax kernel parity tests on the OpenCL CPU device pass (MiniMax's is not shipped). - Installed from the archive in fresh Ubuntu 24.04 and 26.04 and Fedora 44 containers: checksums,
setup, the libraries
resolve, the CPU-device parity tests pass, the server answers against its mock engine. - Server tests: 959 passed, 23 skipped.
- On GPUs: the release candidate built from the parent commit (the same source but for these documents), installed from its archive with the PyPI run-time, ran DeepSeek-V4.1-Flash Coder IQ2_XXS on Arc Pro B70 + B65 + B60 and matched the llama.cpp reference at 16 of 16 forced tokens on both saved prompts (no kernel message). GLM-5.3 and Qwen3.8 were not run with the release build: their
numbers above are from development builds.
Known limitations
- Reference agreement is recorded for the files and prompts above only. Other quantisations and long contexts are
not qualified. - MiniMax-M3.1-Flash is not in this release: its code is in the source and is held back (slow, and one development
run ended in a GPU engine reset of unknown cause). - DeepSeek-V4.1's tool calls are read by the server but were not exercised with the model; its thinking is not
qualified. - Qwen3.8 and GLM-5.3 were not rechecked on GPUs after the last shared expert-tier changes.
- A GPU on a slow or shared link can be starved by long expert streaming.
- Experimental and unsupported in this release: the hybrid segment (a model's last blocks on another machine), the
Windows build with the CUDA target (compiles, never run), NVIDIA and AMD GPUs, the inherited deb/rpm packaging and
setup.pyinstaller.
Files
| File | |
|---|---|
xemaya-0.1.0-alpha.1-linux-x86_64.tar.xz |
The engine built for Intel GPUs, the server, the xemaya.sh command, and the complete source of the build |
xemaya-0.1.0-alpha.1-source.tar.gz |
The source alone |
SHA256SUMS |
Checksums of both |
Licence: LGPL-3.0-or-later, with MIT code from Strata, Project Maya and ggml / llama.cpp
(NOTICE).
Maintainer: Jumbomuffin777.
The published build itself repeated the GPU check after these notes were written: 16 of 16 forced tokens on both DeepSeek-V4.1 references on Arc Pro B70 + B65 + B60, no kernel message.
1c3dfd886959f3783fb92d909dbb2703ccc55d0abdaa23a3be8e368d60d8b37a xemaya-0.1.0-alpha.1-linux-x86_64.tar.xz
f6a4f8911c4e4738772ae49f8bbb5eb5d377373293e08f741d60f7d3d82ff890 xemaya-0.1.0-alpha.1-source.tar.gz