Run open-weight AI models on your PC, checked byte for byte.

A seed is a JSON file of 451 to 528 bytes (our 35 seeds, shipped with player 1.3.1 on 2026-09-15). It names the model file on Hugging Face by its SHA-256 (for a model split into several files: a list of the parts).

The player downloads what the seed names, checks every byte and runs the model on your CPU, or refuses.

Download the player

Linux · Windows 11 · command line · free of charge · 22.6 to 42.7 MiB

What's in a seed file?

{"matiere":{"nom":"Qwen3-0.6B-Q8_0.gguf","octets":639446688,"sha256":"9465e63a22add5354d9bb4b99e90117043c7124007664907259bd16d043bb031","substrat":["hf:Qwen/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf"]},"nom":"qwen3-06b","notes":"licence declaree par l hebergeur : apache-2.0. les poids ne sont pas ici : ils sont NOMMES.","recette":{"classe":"x86-64-fils8","empreinte":"869bef8788166262e4a72fe04f0809756f561a2bae03a52a95e3fb7f294b8226"},"scellee_le":"2026-08-17","v":1}

No weights inside. It names the model file: name, size, SHA-256, where to fetch it.

Before you download

Linux x86-64, glibc 2.35 or newer (measured up to 2.44)
Windows 11 x86-64 (builds 26100 and 26200 measured)
Processor with AVX2, FMA and F16C
macOS, ARM, phones; Windows 10 not tested

CPU only · not code-signed · messages in French · free of charge · player source code not published

26 of 35 seeds give the same tokens on Linux and Windows 26 same · 2 differ by OS · 5 not compared · 2 refused Our test questions (1 to 3 per seed); mostly 1 Linux and 1 Windows machine per seed, both Intel, player 1.2 (same engines as 1.3.1); CPU only Measured 2026-09-14 to 2026-09-15 · Checked outside ZMLabs: 0 · Measurements

Seeds · 35 · smallest first

Model file size (the player refuses one larger than your RAM):

If the player accepts a seed, it can still be slow. Start-up was tested when each seed was made, 2026-08-17 to 2026-08-19.

Jump to a size group

Up to 8 GiB · 20

If the player accepts a seed, it can still be slow. Start-up was tested when each seed was made, 2026-08-17 to 2026-08-19.

8 to 16 GiB · 6

Qwen3-8B-Q8_0
qwen3-8b-source · model by Qwen
Passed start-up
Model file
8.11 GiB
Linux and Windows
Same
Licence, as declared
Apache-2.0
DeepSeek-V2-Lite-Chat-Q4_K_M
deepseek-v2-lite · model by deepseek-ai
Passed start-up
Model file
9.65 GiB
Linux and Windows
Same
Licence, as declared
DeepSeek License 1.0 Extra terms
gpt-oss-20b-MXFP4
gpt-oss-20b · model by openai
Refused at start-up
Model file
11.28 GiB
Cause of refusal
The engine loaded another model's chat template.
Licence, as declared
Apache-2.0 Separate usage policy
Why (gpt-oss-20b)
ERNIE-4.5-21B-A3B-Thinking-Q4_K_M
ernie-45-21b-a3b · model by baidu
Passed start-up
Model file
12.42 GiB
Linux and Windows
Differs by OS
Licence, as declared
Apache-2.0
gemma-4-26B_q4_0-it
gemma4-26b-a4b · model by google
Passed start-up
Model file
13.45 GiB
Linux and Windows
Same
Licence, as declared
Apache-2.0 Separate usage policy
Qwen3.6-27B-Q4_K_M
qwen36-27b · model by Qwen
Passed start-up
Model file
15.66 GiB
Linux and Windows
Same
Licence, as declared
Apache-2.0

16 to 32 GiB · 3

hermes-4_3_36b-Q4_K_M
hermes-43-36b · model by NousResearch
Passed start-up
Model file
20.27 GiB
Linux and Windows
Same
Licence, as declared
Apache-2.0 Extra terms not checked
Seed-OSS-36B-Instruct-Q4_K_M
seed-oss-36b · model by ByteDance-Seed
Passed start-up
Model file
20.27 GiB
Linux and Windows
Same
Licence, as declared
Apache-2.0
Qwen3.6-35B-A3B-UD-Q4_K_M
qwen36-35b-a3b · model by Qwen
Passed start-up
Model file
20.61 GiB
Linux and Windows
Differs by OS
Licence, as declared
Apache-2.0

Over 32 GiB · 6

These need a machine with that much RAM, usually a server; on a smaller one the player refuses before downloading. Not run in our measurements from 2026-09-14 to 2026-09-15; start-up tested when the seeds were made, 2026-08-17 to 2026-08-19.

DeepSeek-V4-Flash-UD-Q2_K_XL
deepseek-v4-flash · model by deepseek-ai
Passed start-up
Model file
90.18 GiB in 3 parts
Linux and Windows
Not compared
Licence, as declared
MIT
Hy3-Q2_K_XL
hy3 · model by tencent
Passed start-up
Model file
93.26 GiB
Linux and Windows
Not compared
Licence, as declared
Apache-2.0
MiniMax-M3-UD-Q2_K_XL
minimax-m3 · model by MiniMaxAI
Refused at start-up
Model file
133.15 GiB in 4 parts
Cause of refusal
The engine did not find in the model file a setting it expects (exit code 1).
Licence, as declared
MiniMax Community License Extra terms
Why (minimax-m3)
GLM-5.2-UD-Q2_K_XL
glm-52 · model by zai-org
Passed start-up
Model file
236.44 GiB in 7 parts
Linux and Windows
Not compared
Licence, as declared
MIT
Kimi-K2.7-Code-UD-Q2_K_XL
kimi-k27-code · model by moonshotai
Passed start-up
Model file
316.15 GiB in 8 parts
Linux and Windows
Not compared
Licence, as declared
Modified MIT License (Kimi K2.7 Code) Extra terms
Kimi-K3-UD-Q2_K_XL
kimi-k3 · model by moonshotai
Passed start-up
Model file
802.13 GiB in 19 parts
Linux and Windows
Not compared
Licence, as declared
Kimi K3 License Extra terms

Licences as declared by each publisher, read 2026-08-07. Not legal advice.

Full table, in French · Catalogue data (JSON, 237.7 KiB) · The task scores in this JSON are for other quantizations, not these files.