Measurements
Compare your answer with ours
- Seed
qwen3-06b· 464 bytes, names Qwen3-0.6B-Q8_0.gguf · 639,446,688 bytes- Question
Quelle est la capitale de la France ?
Type it exactly, in French. It means: “What is the capital of France?”- When
- Ask it first, right after
germer, in a new--maisonfolder. - Settings
- temperature 0 · top-k 1 · sampling seed 1 · up to 256 tokens · no prompt cache
- Expected
- 12 tokens · SHA-256 of the token IDs
72ede708a51254a736c03ecc883d0656d0f2e382e2aa288f33f5104663e402cfClick once to select the whole value. - Text
- La capitale de la France est **Paris**.
- Measured
- 31 runs · 12 machines (11 Linux, 1 Windows 11 VM) · AMD 6, Intel 5, not recorded 1 · all rented · to · 0 differences · as of
- Linux, processor AMD EPYC 7713, rented, : same tokens
- Linux, processor not recorded, rented, : same tokens
- Linux, processor AMD EPYC 7K62, rented, : same tokens
- Linux, processor Intel Xeon E5-2673 v4, rented, : same tokens
- Linux, processor AMD EPYC 7452, rented, : same tokens
- Linux, processor AMD Ryzen 5 5600X, rented, : same tokens
- Linux, processor Intel Xeon Gold 6342, rented, : same tokens
- Linux, processor Intel Xeon E5-2690 v4, rented, : same tokens
- Windows 11, processor Intel Xeon E5-2673 v4, rented, to : same tokens
- Linux, processor AMD EPYC 7551P, rented, : same tokens
- Linux, processor Intel Xeon E5-2680 v4, rented, : same tokens
- Linux, processor AMD Ryzen 7 9800X3D, rented, : same tokens
- Outside ZMLabs
- Checked outside ZMLabs: 0 (as of )
This fingerprint identifies the answer, not the model: qwen3-4b, qwen3-8b and qwen3-8b-source give the same tokens to this question. The byte check on download identifies the model file.
Measured
- Claim
- Same 12 tokens for the test question of
qwen3-06b - Scope
- 31 runs · 12 machines (11 Linux, 1 Windows 11 VM) · AMD and Intel · all rented · 0 differences · as of
- Date
- to
- Claim
- 26 of 35 seeds give the same tokens on Linux and Windows
- Scope
- 2 differ by OS · 5 not compared · 2 refused · 1 to 3 test questions · mostly 2 machines per seed, both Intel, player 1.2
- Date
- to
- Claim
- 20 of 35 give the same tokens on AMD and Intel
- Scope
- 13 not measured · 2 refused · same OS only · 2 to 10 machines
- Date
- to
- Claim
- 273 expected answers on 28 seeds
- Scope
- Linux and Windows differ on 4 test questions for 2 seeds (2 first differences, on 1 Linux and 1 Windows machine; later questions of the same conversation carry them)
- Date
- to
- Claim
- 200 answers from earlier conversations of 20 seeds obtained again byte for byte, 681 times in all
- Scope
- 0 differences · 615 of the 681 on a machine where that answer had not been obtained, 106 of them on Windows for answers obtained on Linux · 11 machines ·
llama31-8b: run again with its birth date given; resumed without it, refused (its chat template includes the date) - Date
- to
- Claim
- 33 of 35 seeds passed start-up
- Scope
- 2 refused at start-up, cause published · tested when each seed was made (Linux, glibc 2.35) · the 6 largest were not run in our measurements from to
- Date
- to
- Data
catalogue.json; date from our run log (SHA-256 published)
- Claim
- 4 of the 6 largest seeds ran a conversation of 10 exchanges again, byte for byte
- Scope
- in a second process, on the original rented server · our unreleased working build, not player 1.3.1 ·
kimi-k3: only a variant computing on 48 threads, not the catalogue seed ·minimax-m3refused at start-up, cause on its page - Date
- Claim
deepseek-v4-flashstarted and answered with player 1.3.1 as served- Scope
- one rented AMD EPYC 7V12 server · 90.18 GiB of model files copied in beforehand, checked byte for byte · start-up passed in 130 s · 3.3 to 3.4 tokens/s · same tokens as on for 2 questions · Linux only; AMD vs Intel not compared with the player
- Date
- Claim
- 65 refusals in our own tests, each with its cause
- Scope
- forged or edited conversation files · model files other than the one the seed names · missing or wrong dates · a dead proxy standing in for Hugging Face · a start-up timer forced to 1 s · a model file outside the folder · a
--matierepath that does not exist · allama31-8bconversation resumed later on another server (its chat template includes the date) - Date
- to
- Claim
- The forged conversation file
journal_forge.jsonis refused - Scope
- 11 runs, all refused · a setting the player does not allow (
prompt) · exit code 1, no answer produced · 8 machines, Linux and Windows 11 · players 1.1 to 1.3.1 - Date
- to
- Claim
- Speed ranges for 28 seeds
- Scope
- 3 rented AMD EPYC servers shared with other work · 8 threads · short answers in a new conversation · player 1.2, same engines · partial runs not counted:
gemma4-12b - Date
- Data
vitesse_mesuree.json(derived copy)
- Claim
- Peak memory for 2 seeds
- Scope
qwen3-06b909 MiB,llama32-1b1.33 GiB · 1 rented Linux server · largest process only, not the player and its engine together · one run · player 1.3.1- Date
- Data
parcours.json
- Claim
- Same .mem file on Linux and Windows 11: 3,729 bytes, same cover fingerprint
- Scope
rangeron allama32-1bconversation · coverf0d4261ee8e3d4c8…on both · Intel Xeon E5-2680 v4 (Linux) and Intel Xeon Platinum 8272CL (Windows 11), both Intel: makers not compared ·verifierpassed for 2 of 2 seeds on each · player 1.3.1 · not for the 5 seeds whose model comes in several parts: the player's code refuses them (tested, never run on these seeds)- Date
- Claim
- 3 of 3 answers checked again byte for byte with Hugging Face cut off
- Scope
llama32-1b· player 1.3.1 · new container on a rented server, Debian 13 (trixie), AMD Ryzen 7 9800X3D · after download, the addresses of Hugging Face pointed to the machine itself; other network links were not cut- Date
- Claim
- Same 12 tokens for the test question of
qwen3-06bwith glibc 2.44 - Scope
- Arch Linux, new container on a rented server, AMD Ryzen 9 9900X · player 1.3.1
- Date
- Claim
- Archives 1.3.1
- Scope
- 2 files, SHA-256 published, not signed
- Date
- Claim
- This site's commands, run as shown
- Scope
- Ubuntu 22.04.5 LTS, new container on a rented server · Windows 11 VM used before, its earlier files moved aside · rented
- Date
- Data
parcours.json
Not measured
- Checked by people outside ZMLabs: 0 (as of )
- macOS, ARM, phones: no build
- GPU: the player uses the CPU
- Windows 10: not tested
- Speed on your computer
- Peak memory for 33 seeds
- Dense models of 7 to 8 billion parameters on a laptop: never timed
- Answer quality: not judged. The player only checks that an answer is not degenerate. Task scores in the catalogue data are for other quantizations, not these files.
- Linux vs Windows for 5 seeds; AMD vs Intel for 13
- The 2 refused seeds, run again with player 1.3.1 (refused in August; same engine build)
- SmartScreen after a browser download (as of )
- A CPU without AVX2 (checked by the player, never run)
- What happens with
--quand-meme(it skips the RAM and processor checks)
Outside the player: internal measurements, code not published
- Claim
- Same bytes on graphics cards from NVIDIA and AMD, for a pilot model
- Scope
- ternary model of 0.6 billion parameters, not a catalogue seed · 20 of 20 answers and 640 of 640 logit fingerprints identical to a CPU reference, on each card · NVIDIA GeForce RTX 4070 Laptop and AMD Radeon 780M: our laptop, Windows 11, D3D12 · NVIDIA Tesla T4: rented, Ubuntu 22.04.5 LTS, Vulkan · 1.8 to 3.2 tokens/s · no Apple or Intel graphics card tested · outside the player, which uses the CPU
- Date
- to
- Data
hors_joueur.json
- Claim
deepseek-v4-flash(90.18 GiB) checked from a laptop that does not hold its weights- Scope
- Windows 11 laptop, 15.23 GiB of RAM · 3 of 3 answers and computation fingerprints (Merkle roots) identical on two rented servers, AMD EPYC 7V12 and Intel Xeon Gold 6132 · a forged step located in 14 comparisons and recomputed on the laptop in 0.3 s; the laptop received that step's weights (592.0 MiB), not the model's · measurement only, no service offered · tools not published
- Date
- Data
hors_joueur.json
- Claim
gpt-oss-120b(59.03 GiB), not a catalogue seed, answering with its weights read from disk as needed- Scope
- two rented servers: 7.3 then 5.2 tokens/s, peak memory up to 23.42 GiB; the second used the model file as republished upstream · our Windows 11 laptop, Linux subsystem limited to 10.7 GiB: 2.3 tokens/s, peak memory 9.30 GiB · 19 to 20 of 20 programming tasks solved, checked by running their tests · answers differ from one machine to another, at most 4 of 20 identical · speed depends on the disk and the server's load
- Date
- to
- Data
hors_joueur.json
Machines
| ID | Host | System | CPU | Maker | Rented or owned | Player | Date |
|---|---|---|---|---|---|---|---|
M1 | H1 | Linux | AMD EPYC 7713 | AMD | rented | unreleased working build | |
M2 | H2 | Linux | AMD EPYC 9654 | AMD | rented | unreleased working build | |
M3 | H3 | Linux | not recorded | not recorded | rented | unreleased working build | |
M4 | H4 | Linux | AMD EPYC 7K62 | AMD | rented | unreleased working build | |
M5 | H5 | Linux | Intel Xeon E5-2673 v4 | Intel | rented | unreleased working build | |
M6 | H6 | Linux | AMD EPYC 7452 | AMD | rented | 1.1 | |
M7 | H7 | Linux | AMD Ryzen 5 5600X | AMD | rented | 1.1 | |
M8 | H8 | Linux | Intel Xeon Gold 6342 | Intel | rented | 1.1 | |
M9 | H9 | Windows 11 Home | AMD Ryzen 7 8845HS | AMD | owned | 1.1 | |
M10 | H10 | Linux | Intel Xeon E5-2690 v4 | Intel | rented | 1.1 | |
M11 | H11 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.1 | |
M12 | H12 | Linux | AMD EPYC 7K62 | AMD | rented | 1.2 | |
M13 | H13 | Linux | AMD EPYC 7551P | AMD | rented | 1.2 | |
M14 | H11 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.2 | |
M15 | H14 | Linux | Intel Core i7-6700 | Intel | rented | 1.2 | |
M16 | H11 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.2 | |
M17 | H15 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.2 | |
M18 | H16 | Debian GNU/Linux 12 (bookworm) | Intel Xeon E5-2680 v4 | Intel | rented | 1.3 | |
M19 | H11 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.3 | |
M20 | H17 | Debian GNU/Linux 13 (trixie) | AMD Ryzen 7 9800X3D | AMD | rented | 1.3.1 | |
M21 | H11 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.3.1 | |
M22 | H18 | Debian 12 (bookworm) | Intel Xeon E5-2680 v4 | Intel | rented | 1.2 | |
M23 | H19 | Fedora Linux 42 (Container Image) | AMD EPYC 7551 | AMD | rented | 1.2 | |
M24 | H20 | Windows 11 Pro | Intel Xeon Platinum 8171M | Intel | rented | 1.2 | |
M25 | H21 | openSUSE Leap 15.6 | AMD EPYC 7402P | AMD | rented | 1.3 | |
M26 | H22 | Windows 11 Pro | Intel Xeon E5-2673 v4 | Intel | rented | 1.3 |
Engines and seeds are identical across players 1.1 to 1.3.1.
identites_os.json is the exact table inside player 1.3.1; its machine labels are internal names.
Each rented Linux server counts as one machine: from one rental to the next, we did not check whether the physical host was the same.
First differences between Linux and Windows (2)
ernie-45-21b-a3b· Quelle est la capitale de la France ? (“What is the capital of France?”)Linux:
67a5d4a7…f993f· Windows 11:7535bf8d…3fe1fLater questions that carry it: What is the capital of Italy?, Continue la suite : 2, 4, 8, 16, (“Continue the sequence: 2, 4, 8, 16,”)
qwen36-35b-a3b· Continue la suite : 2, 4, 8, 16, (“Continue the sequence: 2, 4, 8, 16,”)Linux:
f40f42e8…8e4a5· Windows 11:527d5aab…1e907
Data from our run logs: 6 JSON files, 2 test vectors
| File | Size | SHA-256 |
|---|---|---|
catalogue.json | 237.7 KiB | 3dfa0d66…7dba6 |
identites_os.json | 24.6 KiB | 30528bfd…38b97 |
quittances_attendues.json | 607.1 KiB | 5ecaeeb7…d0153 |
vitesse_mesuree.json | 67.4 KiB | 71fe5cc5…c4bed |
mesures_complementaires.json | 22.1 KiB | fe2d3aef…1ec3f |
hors_joueur.json | 17.5 KiB | a482cb0e…139b9 |
journal_A.json (negative test vector) | 1.1 KiB | 86192dbb…e419c |
journal_forge.json (negative test vector) | 1.3 KiB | ce0cff42…820a7 |