Meta-Llama-3.1-8B-Instruct-Q4_K_M

This seed file names one model file by its SHA-256. The player downloads it, checks every byte and runs it on your CPU, or refuses.

Passed start-up
Same on Linux and Windows
Same on AMD and Intel, same OS

Use this seed

Download the player Have the player? Run it

Needs

RAM: the player refuses below 4.58 GiB
Peak memory: not measured for this seed. The player compares your RAM with the file size, not with the memory it will use; on the 2 seeds we measured, the largest process peaked 49% and 77% above the file size (Linux, 2026-09-23)
Free disk ≥ 4.81 GiB
x86-64 processor with AVX2, FMA and F16C
Linux (glibc 2.35 or newer) or Windows 11 (builds 26100 and 26200 measured)

The player checks RAM, free disk and the processor before downloading, and stops if any of them falls short; a value it cannot read does not stop it.

Downloads 4.58 GiB from Hugging Face.

Seed file · 521 bytes · already in the player's archive, catalogue folder

Seed llama31-8b · 521 bytes

Model by meta-llama · file converted by bartowski (bartowski/Meta-Llama-3.1-8B-Instruct-GGUF)

Llama 3.1 Community License · Extra terms · licence text read on the publisher's public mirror · as declared, read 2026-08-07 · not legal advice

Linux vs Windows comparison: same tokens on 3 test questions, 2 machines, 2026-09-15.

Run it

From the unpacked player folder. --maison is the folder for this model and your conversations.

Linux · bash

# germer: download the model file and check every byte./memown germer catalogue/llama31-8b.graine.json --maison ~/ia/llama31-8b

Expected: a line starting with VERDICT : VERT (VERT: all four checks passed, including the byte-for-byte check)

Windows 11 · PowerShell

# germer: download the model file and check every byte.\memown.exe germer catalogue\llama31-8b.graine.json --maison "$env:USERPROFILE\ia\llama31-8b"

Expected: a line starting with VERDICT : VERT (VERT: all four checks passed, including the byte-for-byte check)

Compare your answer with ours

Ask it first, right after germer, in a new --maison folder. Type the question exactly, in French: the published fingerprint is for these exact bytes.

Expected: 10 tokens, seen on 10 machines (Linux and Windows 11), 2026-09-07 to 2026-09-15, 0 differences · fingerprint de80b4801761df28c6ddab7880167707da015376a53d2a31af77a8037b647914 Click once to select the whole value.

Linux · bash

# parler: ask a question./memown parler --maison ~/ia/llama31-8b --dire "Quelle est la capitale de la France ?"

grep -c '"ids_sha256":"de80b4801761df28c6ddab7880167707da015376a53d2a31af77a8037b647914"' ~/ia/llama31-8b/vies/vie/journal.json

Expected: 1

Windows 11 · PowerShell

# parler: ask a question.\memown.exe parler --maison "$env:USERPROFILE\ia\llama31-8b" --dire "Quelle est la capitale de la France ?"

Select-String -Quiet -SimpleMatch -Path "$env:USERPROFILE\ia\llama31-8b\vies\vie\journal.json" -Pattern '"ids_sha256":"de80b4801761df28c6ddab7880167707da015376a53d2a31af77a8037b647914"'

Expected: True

1 or True: your tokens are the same as ours. 0 or False: they differ. Tell us either way.

To ask your own question, change the text after --dire. Keep the test question to compare with ours.

This fingerprint identifies the answer, not the model: llama32-1b and llama32-3b give the same tokens to this question. The byte check on download identifies the model file.

This model's chat template includes the date. The published answer was recorded on 2026-09-07, 2026-09-14 and 2026-09-15; on another date it may differ.

What this seed names

Model file: Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
Size: 4,920,739,232 bytes (4.58 GiB)
SHA-256: 7b064f5842bf9532c91456deda288a1b672397a54fa729aa665952863033557c
Hosted at: Hugging Face gave the same SHA-256 for this file (checked 2026-09-23) · If that file changes, the player refuses it.
Licence: Llama 3.1 Community License · Extra terms
Seed file: llama31-8b.graine.json · 521 bytes · cc8004821c676321e0a63f077d4493560fb8d0a0a7387ec7f33bab05ffe5b301

What we measured

Start-up: passed on 2026-08-17 when the seed was made (Linux, glibc 2.35).
Linux vs Windows: same tokens · 3 test questions · 2 machines · 2026-09-15 · data (JSON)
AMD vs Intel, same OS: same tokens · 3 test questions (1 to 3 per machine) · 4 machines · 2026-09-14 to 2026-09-15
16 expected answers · up to 6 machines · 0 differences
8.8–19.1 tokens/s · 3 rented AMD EPYC servers shared with other work · 8 threads · short answers in a new conversation · player 1.2, same engines as 1.3.1 · 2026-09-14. Your speed: not measured.

Not measured for this seed

Peak memory

Not measured for any seed: see Measurements.

16 expected answers

Question	Expected	Measured
`Quelle est la capitale de la France ?` (“What is the capital of France?”)	10 tokens · SHA-256 of the token IDs `de80b480…47914`	6 machines
`Quelle est la capitale de la France ?` (“What is the capital of France?”)	10 tokens · SHA-256 of the token IDs `de80b480…47914`	5 machines
`Quelle est la capitale de la France ?` (“What is the capital of France?”)	10 tokens · SHA-256 of the token IDs `de80b480…47914`	2 machines
`Continue la suite : 2, 4, 8, 16,` (“Continue the sequence: 2, 4, 8, 16,”)	49 tokens · SHA-256 of the token IDs `c49cd647…659ce`	6 machines
`What is the capital of Italy?`	8 tokens · SHA-256 of the token IDs `4468ec6e…24ac1`	2 machines
`What is the capital of Italy?`	8 tokens · SHA-256 of the token IDs `4468ec6e…24ac1`	2 machines
`Continue la suite : 2, 4, 8, 16,` (“Continue the sequence: 2, 4, 8, 16,”)	32 tokens · SHA-256 of the token IDs `2909d31a…c7f77`	2 machines
`Continue la suite : 2, 4, 8, 16,` (“Continue the sequence: 2, 4, 8, 16,”)	32 tokens · SHA-256 of the token IDs `2909d31a…c7f77`	2 machines
`Explique en deux phrases pourquoi le ciel est bleu.` (“Explain in two sentences why the sky is blue.”)	92 tokens · SHA-256 of the token IDs `cdde3ead…07e0c`	6 machines
`Recopie exactement : le chat dort sur le tapis rouge et la pluie tombe doucement.` (“Copy exactly: the cat sleeps on the red rug and the rain falls gently.”)	19 tokens · SHA-256 of the token IDs `d3d99286…f212b`	6 machines
`Donne trois nombres premiers entre 10 et 30.` (“Give three prime numbers between 10 and 30.”)	25 tokens · SHA-256 of the token IDs `7dfab34d…52e59`	6 machines
`Traduis en anglais : la mer est calme ce soir.` (“Translate into English: the sea is calm tonight.”)	7 tokens · SHA-256 of the token IDs `52654b43…29f8d`	6 machines
`Quel est le plus grand océan du monde ?` (“What is the largest ocean in the world?”)	34 tokens · SHA-256 of the token IDs `e9c19ec1…82656`	6 machines
`Écris une phrase de politesse pour finir un e-mail.` (“Write a polite sentence to end an email.”)	27 tokens · SHA-256 of the token IDs `8f5f7974…8ad7a`	6 machines
`Résume en une phrase ce que fait un four à pain.` (“Sum up in one sentence what a bread oven does.”)	36 tokens · SHA-256 of the token IDs `c7b24590…7c97d`	6 machines
`Continue la suite : 3, 6, 12, 24,` (“Continue the sequence: 3, 6, 12, 24,”)	46 tokens · SHA-256 of the token IDs `eacd8d92…387dc`	6 machines

The seed file (JSON, 521 bytes)

{"matiere":{"nom":"Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf","octets":4920739232,"sha256":"7b064f5842bf9532c91456deda288a1b672397a54fa729aa665952863033557c","substrat":["hf:bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf"]},"nom":"llama31-8b","notes":"licence declaree par l hebergeur : llama3.1. les poids ne sont pas ici : ils sont NOMMES.","recette":{"classe":"x86-64-fils8","empreinte":"869bef8788166262e4a72fe04f0809756f561a2bae03a52a95e3fb7f294b8226"},"scellee_le":"2026-08-17","v":1}

Licence as we read it, in French

Trois obligations SUIVENT la redistribution : joindre une copie de l'accord ; afficher « Built with Llama » en évidence sur le site, l'interface, un billet, la page « à propos » ou la documentation ; conserver un fichier « Notice » portant « Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc.
All Rights Reserved. ».
Tout modèle entraîné ou amélioré à partir de ces matières et distribué doit porter « Llama » AU DÉBUT DE SON NOM.
Seuil : au-delà de « 700 million monthly active users » au mois précédent, il faut demander une licence à Meta.
La politique d'usage acceptable est incorporée par référence. (Version du 23 juillet 2024.)
texte de la licence lu au miroir public de l'éditeur — la page de l'hébergeur exige une authentification (HTTP 401) ; l'identité octet à octet des deux copies n'est pas prouvée

github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE

Full record, in French