Llama-3.1-405B

meta-llama Llama 3.1 Community License out 2 years ago

406B parameters, text in, text out.

Verified 2026-10-25 Rank #204 in LLMs · score 88.5 method v1 Sources 3 Downloads 30d 34.9k · #412 of 619 tracked

Can I run it?244GB at INT4Estimateneeds a multi-GPU nodemethodCommunity: Show HN: Run Llama3.1 405B on a 8GB VRAM challenge [video] →

Can I use it?Llama 3.1 Community License'Built with Llama' attribution; >700M MAU clauselicence file2026-10-24

Do people use it?34.9kdownloads, 30 days#412 of 619 we track · below the median+1 likes, downloads −19% this week992 likesHugging Face2026-10-24

Worth reading

Picked by rule from 54 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 427 HN pts

Llama 3.1 405B now runs at 969 tokens/s on Cerebras Inference

Cerebras Inference runs Llama 3.1 405B at 969 tokens per second, achieving the highest recorded performance at 128K context length.

cerebras.ai 2024-11-19 427 HN pts · 156 comments 553 words

Latest · 58 days ago

Running Llama 3.1 405B

Llama 3.1 405b base is no longer hosted by any public inference provider.

a11ce.com 2026-08-28 4 HN pts 428 words

25 best of 54 tracked

The 2 picks above, plus these 23. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

CodeShow HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPUgithub.com · 2026-02-21
ArticleLlama 3.1llama.meta.com · 2024-07-23 · 437 HN pts · official
Codeg1: Using Llama-3.1 70B on Groq to create o1-like reasoning chainsgithub.com · 2024-09-15
CodeLlama 3.1 Omni Modelgithub.com · 2024-09-18
ArticleMeta's Llama 3.1 can recall 42 percent of the first Harry Potter bookunderstandingai.org · 2025-06-15 · 191 HN pts
CodeShow HN: Tune LLaMa3.1 on Google Cloud TPUsgithub.com · 2024-09-11 · 189 HN pts
CodeLlama 3.1 in Cgithub.com · 2024-07-24 · 212 HN pts
CodeLongwriter – Increase llama3.1 output to 10k wordsgithub.com · 2024-10-07 · 154 HN pts
ArticleHermes 3: The First Fine-Tuned Llama 3.1 405B Modellambdalabs.com · 2024-08-15 · 146 HN pts
ArticleLlama 3.1: Our most capable models to dateai.meta.com · 2024-07-23 · 27 HN pts · official
DiscussionCerebras launches inference for Llama 3.1; benchmarked at 1846 tokens/s on 8Btwitter.com · 2024-08-27 · 95 HN pts
ArticleWhy Llama 3.1 is Importanttimkellogg.me · 2024-07-23
ArticleNvidia presents Llama-Mesh: Generating 3D Mesh with Llama 3.1 8Bresearch.nvidia.com · 2024-11-17
ArticleMeta Llama 3.1 405Bopenrouter.ai · 2024-07-23
ArticleCoding with Llama 3.1, New DeepSeek Coder and Mistral Largeaider.chat · 2024-07-25
ArticleLlama-3.1 supports tool calls via promptingbraintrust.dev · 2024-07-26
ArticleMeta releases an open-weights GPT-4-level AI model, Llama 3.1 405Barstechnica.com · 2024-07-23
ArticleThe Future of AI: Synthetic Data Gen with Llama 3.1 405B and Rafttechcommunity.microsoft.com · 2024-09-17
Codellama31.md at main · huggingface/bloggithub.com · undated
ArticleShow HN: Run Llama 3.1 8B in the browserapp.wiz.chat · 2024-07-28
VideoRunning the latest Llama 3.1 8B on Raspberry Pi [video]youtube.com · 2024-07-30
VideoShow HN: Run Llama3.1 405B on a 8GB VRAM challenge [video]youtube.com · 2024-08-01
ArticleShow HN: Finetune Llama-3.1 2x faster in a Colabcolab.research.google.com · 2024-07-23

Show all 23

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 11 quantized builds and 100 Spaces.

As a service0no API provider listed
On own hardware11quantized builds
Built on2017 fine-tunes · 3 adapters · 0 merges
In apps100+Hugging Face Spaces

Top Spaces running it: xlr8harder/completion ♥ 21 · xlr8harder/sydney ♥ 15 · artificialguybr/CanIRunThisLLM ♥ 9 · Babu07/meta-llama-Meta-Llama-3.1-405B ♥ 8

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Needs an 8×80 GB node even at INT4

About 244 GB at INT4, 487 GB at FP8, 974 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

10 dated articles and 0 repo commits so far

Teal is the weights (2024-07-16), amber triangles are 10 dated articles (each links to it).

What people built on it

31 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

31 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF ThomasBaruzier/Meta-Llama-3.1-405B-Instruct-GGUF for llama.cpp, Ollama, LM Studio 46.5k downloads FP8 NousResearch/Hermes-4-405B-FP8 for vLLM, SGLang 9.9k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Facts

17 fields

Listed in LLMs

Facts

Llama-3.1-405B has 406B parameters. (safetensors metadata, observed 2026-10-24)
License: Llama 3.1 Community License (friction: attribution). 'Built with Llama' attribution; >700M MAU clause. (license file, observed 2026-10-24)
Created 2024-07-16, last modified 2024-09-25. (HF API, observed 2026-10-24)
Access is gated on Hugging Face (manual). (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	LlamaForCausalLM
Model type	llama
Params by dtype	BF16: 405,853,388,800
Size band	XXL
MoE	no
Gated	manual
Papers	arXiv:2204.05149
License	Llama 3.1 Community License (llama3.1)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · LlamaForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · LlamaForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Llama-3.1-405B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	974 GB · Cluster (≥640 GB)
FP8 (1 byte/param)	487 GB · Multi-GPU node (<640 GB)
INT4 (0.5 bytes/param)	244 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

31 variants

Lineage

Original model Family: meta-llama/Llama-3.1 31 derivatives on the Hub

quantized: 11finetune: 17adapter: 3

Top variants (31 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
ThomasBaruzier/Meta-Llama-3.1-405B-Instruct-GGUF	quantized	GGUF	—	—	llama3.1	ThomasBaruzier	2	46.5k
NousResearch/Hermes-3-Llama-3.1-405B-FP8	finetune	-FP8	406B	XXL	llama3	NousResearch	29	29.3k
hugging-quants/Meta-Llama-3.1-405B-BNB-NF4-BF16	quantized	-BNB	117B	XXL	llama3.1	hugging-quants	2	21.6k
NousResearch/Hermes-4-405B	finetune	—	406B	XXL	llama3	NousResearch	97	8k
NousResearch/Hermes-4-405B-FP8	quantized	-FP8	406B	XXL	llama3	NousResearch	38	9.9k
Bllossom/llama-3.1-Korean-Bllossom-405B	quantized	—	410B	XXL	llama3.1	Bllossom	56	4.1k
allenai/Llama-3.1-Tulu-3-405B-SFT	finetune	—	—	—	llama3.1	allenai	11	6k
RedHatAI/Meta-Llama-3.1-405B-FP8	quantized	-FP8	—	—	llama3.1	RedHatAI	0	5k
deepcogito/cogito-v2-preview-llama-405B	finetune	—	406B	XXL	llama3.1	deepcogito	14	3.3k
clowman/Llama-3.1-405B-Instruct-AWQ-Int4	quantized	AWQ	406B	XXL	llama3.1	clowman	0	2.9k
kgrabko/JiRackTernary_405b	quantized	—	108B	XXL	other	kgrabko	4	2k
hugging-quants/Meta-Llama-3.1-405B-BNB-NF4	quantized	-BNB	418B	XXL	llama3.1	hugging-quants	2	1.7k
Deniss8686/Hermes-4-405B	finetune	—	406B	XXL	llama3	Deniss8686	0	538
genevaquant/Hermes-4-405B-FP8	quantized	-FP8	406B	XXL	llama3	genevaquant	0	469
nick911/hermes-3-llama-3-1-405b-fp8	finetune	-fp8	406B	XXL	llama3	nick911	0	307
nisten/lobotollama-368b-base	finetune	—	368B	XXL	—	nisten	1	233
stapspop/Hermes-3-Llama-3.1-405B	finetune	—	406B	XXL	llama3	stapspop	0	249
genevaquant/Hermes-4-405B	finetune	—	406B	XXL	llama3	genevaquant	0	222
nick911/hermes-3-llama-3-1-405B	finetune	—	406B	XXL	llama3	nick911	0	212
axolotl-ai-co/meta-llama-3.1-405b-alpaca	adapter	—	—	—	llama3.1	axolotl-ai-co	2	90
rhuanmatias/Hermes-4	quantized	—	406B	XXL	llama3	rhuanmatias	1	121
ericflo/Llama-3.1-SyntheticPython-405B-Base-LoRA	adapter	—	—	—	—	ericflo	1	42
gtest23/H3-810B-fixed-attn	finetune	—	807B	XXL	llama3	gtest23	0	85
RJ1/Hermes-3-Llama-3.1-405B	finetune	—	406B	XXL	llama3	RJ1	0	77
Maveth/Hermes-3-Llama-3.1-405B-rk3588-1.1.2	finetune	—	—	—	llama3	Maveth	0	74
indigoskyai/Hermes-3-Llama-3.1-405B	finetune	—	406B	XXL	llama3	indigoskyai	0	58
mradermacher/Meta-Llama-3.1-405B-i1-GGUF	finetune	GGUF	—	—	llama3.1	mradermacher	1	0
Dongz/llama31-405b-quant	finetune	-quant	—	—	—	Dongz	1	0
Tensordyne/Llama-3.1-405B	quantized	—	—	—	llama3.1	Tensordyne	1	0
mradermacher/Meta-Llama-3.1-405B-GGUF	finetune	GGUF	—	—	llama3.1	mradermacher	0	0
ericflo/Llama-3.1-SyntheticPython-405B-Base	adapter	—	—	—	—	ericflo	0	0

Articles and coverage

54 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official3

Llama 3.1

llama.meta.com2024-07-23officialHN front pageabout the Llama 3.1 familyHN 437 pts · 269 comments42 words

Llama 3.1 405B now runs at 969 tokens/s on Cerebras Inference

cerebras.ai2024-11-19HN front pageHN 427 pts · 156 comments553 words

Cerebras Inference runs Llama 3.1 405B at 969 tokens per second, achieving the highest recorded performance at 128K context length. The platform delivers a 240ms time-to-first-token and offers general availability in Q1 2025 at $6 per million input tokens and $12 per million output tokens. This result is 12x faster than the best GPU result and 75x faster than AWS.written, reviewed, cited

Llama 3.1: Our most capable models to date

ai.meta.com2024-07-23officialauthorityabout the Llama 3.1 familyHN 27 pts · 3 comments2,136 words

Analysis6

Running Llama 3.1 405B

a11ce.com2026-08-28HN 4 pts · 0 comments428 words

Llama 3.1 405b base is no longer hosted by any public inference provider. You must run it yourself on an on-demand GPU instance for ~$20/hr, ensuring the instance has >500GB of tmpfs and disabling NCCL_NVLS_ENABLE. The setup deploys a 4x H200 pod to download weights from a Cloudflare R2 bucket and serve the model via vllm.written, reviewed, cited

Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

understandingai.org2025-06-15HN front pageabout the Llama 3.1 familyHN 191 pts · 305 comments

Show HN: Tune LLaMa3.1 on Google Cloud TPUs

github.com2024-09-11authorityHN front pageabout the Llama 3.1 familyHN 189 pts · 52 comments869 words

Llama 3.1 in C

github.com2024-07-24authorityHN front pageabout the Llama 3.1 familyHN 212 pts · 36 comments5,858 words

Longwriter – Increase llama3.1 output to 10k words

github.com2024-10-07authorityHN front pageabout the Llama 3.1 familyHN 154 pts · 29 comments6,696 words

Hermes 3: The First Fine-Tuned Llama 3.1 405B Model

lambdalabs.com2024-08-15HN front pageHN 146 pts · 67 comments530 words

Nous Research launched Hermes 3, the first full-parameter fine-tune of Meta's Llama 3.1 405B model, trained on an 8-node Lambda 1-Click Cluster. The model achieves a 50% reduction in VRAM and disk requirements through Neural Magic's FP8 quantization, enabling single-node operation. Hermes 3, which has been downloaded over 33 million times, is available for free via Lambda's Chat Completions API and Lambda Chat interface.written, reviewed, cited

Community1

Cerebras launches inference for Llama 3.1; benchmarked at 1846 tokens/s on 8B

twitter.com2024-08-27about the Llama 3.1 familyHN 95 pts · 42 comments

All coverage tracked for this model (54) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
llama31.md at main · huggingface/blog	github.com	—	2026-10-07	exa
LLM Deep-dive: Llama 3.1	predli.com	—	2026-10-07	exa
Llama 3.1 405B Review — Benchmarks, Pricing & AI Panel Verdict \| TopReviewed.ai	topreviewed.ai	—	2026-10-07	exa
Llama 3.1 405B: Complete Guide — Architecture, Benchmarks, Fine-Tuning, Deployment & Use Cases 2026 \| aimodels.in	aimodels.in	—	2026-10-07	exa
Llama 3.1: An In-Depth Analysis of the Next Generation Large Language Model	researchgate.net	—	2026-10-07	exa
Running Llama 3.1 405B	a11ce.com	2026-08-28	2026-10-07	hn
Meta Llama 3.1 405B Review — The First Open-Weight Frontier Model — ChatForest	chatforest.com	2026-05-13	2026-10-07	exa
Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPU	github.com	2026-02-21	2026-10-07	hn
Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book	understandingai.org	2025-06-15	2026-10-07	hn
Nvidia's new Llama-3.1 Nemotron Ultra outperforms DeepSeek R1 at half the size	venturebeat.com	2025-04-11	2026-10-07	hn
Llama 3.1 405B now runs at 969 tokens/s on Cerebras Inference	cerebras.ai	2024-11-19	2026-10-07	hn
Nvidia presents Llama-Mesh: Generating 3D Mesh with Llama 3.1 8B	research.nvidia.com	2024-11-17	2026-10-07	hn
Benchmarking Llama 3.1 405B on 8x AMD MI300X GPUs	dstack.ai	2024-10-09	2026-10-07	hn
Longwriter – Increase llama3.1 output to 10k words	github.com	2024-10-07	2026-10-07	hn
Llama 3.1 Omni Model	github.com	2024-09-18	2026-10-07	hn
The Future of AI: Synthetic Data Gen with Llama 3.1 405B and Raft	techcommunity.microsoft.com	2024-09-17	2026-10-07	hn
g1: Using Llama-3.1 70B on Groq to create o1-like reasoning chains	github.com	2024-09-15	2026-10-07	hn
Show HN: Tune LLaMa3.1 on Google Cloud TPUs	github.com	2024-09-11	2026-10-07	hn
Create a free Llama 3.1 405B-powered chatbot on a GitHub repo in <1 min	blog.stephenturner.us	2024-09-11	2026-10-07	hn
Reflection-Llama-3.1-70B is Llama-3 with LoRA	old.reddit.com	2024-09-08	2026-10-07	hn
Cerebras launches inference for Llama 3.1; benchmarked at 1846 tokens/s on 8B	twitter.com	2024-08-27	2026-10-07	hn
Hyperbolic is only provider delivering Llama 3.1 405B BASE model	app.hyperbolic.xyz	2024-08-16	2026-10-07	hn
Hermes 3: The First Fine-Tuned Llama 3.1 405B Model	lambdalabs.com	2024-08-15	2026-10-07	hn
Llama 3.1 In-Depth Analysis: Cutting Through the Noise	blog.spheron.network	2024-08-13	2026-10-07	exa
Open source Claude Artifacts – built with Llama 3.1 405B	github.com	2024-08-04	2026-10-07	hn
Meta's Joe Spisak on Llama 3.1 405B [video]	youtube.com	2024-08-03	2026-10-07	hn
Using Llama3.1 405B to generate political synthetic data	oxen.ai	2024-08-02	2026-10-07	hn
Meta Releases Llama 3.1 405B, Largest Open-Source Model to Date	infoq.com	2024-08-01	2026-10-07	hn
Show HN: Run Llama3.1 405B on a 8GB VRAM challenge [video]	youtube.com	2024-08-01	2026-10-07	hn
Running the latest Llama 3.1 8B on Raspberry Pi [video]	youtube.com	2024-07-30	2026-10-07	hn
Show HN: Run Llama 3.1 8B in the browser	app.wiz.chat	2024-07-28	2026-10-07	hn
Try Llama 3.1 405B on LangCSS (A Tailwind CSS AI)	blog.langcss.com	2024-07-27	2026-10-07	hn
Llama-3.1 supports tool calls via prompting	braintrust.dev	2024-07-26	2026-10-07	hn
CentML gives free access to llama3.1-405B	cserve-llama.centml.com	2024-07-25	2026-10-07	hn
Coding with Llama 3.1, New DeepSeek Coder and Mistral Large	aider.chat	2024-07-25	2026-10-07	hn
Fine-Tuning Llama 3.1 405B on a Single Node	snowflake.com	2024-07-25	2026-10-07	hn
Llama 3.1 405B: Open Model Meets Closed Frontier	jacar.es	2024-07-25	2026-10-07	exa
Llama 3.1 405B, Meta's AI strategy, and the new, open frontier model ecosystem	interconnects.ai	2024-07-24	2026-10-07	hn
Llama 3.1 in C	github.com	2024-07-24	2026-10-07	hn
A Startup Founder Reviews Meta's New Llama 3.1 405B	youtube.com	2024-07-24	2026-10-07	hn
Why Llama 3.1 is Important	timkellogg.me	2024-07-23	2026-10-07	hn
Show HN: Finetune Llama-3.1 2x faster in a Colab	colab.research.google.com	2024-07-23	2026-10-07	hn
Meta releases an open-weights GPT-4-level AI model, Llama 3.1 405B	arstechnica.com	2024-07-23	2026-10-07	hn
Meta Llama 3.1 405B	openrouter.ai	2024-07-23	2026-10-07	hn
Chat with Meta Llama 3.1 405B	llama3.replicate.dev	2024-07-23	2026-10-07	hn
Llama 3.1: 405B, the largest openly available model released	github.com	2024-07-23	2026-10-07	hn
Zuck pod interview: Llama 3.1 405B, Meta's open-source AI beats GPT-4	twitter.com	2024-07-23	2026-10-07	hn
Llama 3.1 405B on Groq	wow.groq.com	2024-07-23	2026-10-07	hn
Llama 3.1: Our most capable models to date	ai.meta.com official	2024-07-23	2026-10-07	hn
Llama 3.1	llama.meta.com official	2024-07-23	2026-10-07	hn
Llama 3.1 405B rumored for release tomorrow (Jul 23)	twitter.com	2024-07-23	2026-10-07	hn
Meta unleashes Llama 3.1 \| UnfoldAI	unfoldai.com	2024-07-23	2026-10-07	exa
Llama 3.1 405B base model available for download	old.reddit.com	2024-07-22	2026-10-07	hn
llama3.1-405B leaked at 3am the day before on 4chan	twitter.com	2024-07-22	2026-10-07	hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink cited in card

David Patterson, Joseph Gonzalez, Urs Hölzle +7 · 2022-04-11 · PDF · HF paper page (13 upvotes)

Similar models

3 models

Similar models

Hermes-3-Llama-3.1-405B · 406B
same size band (XXL), same task (text-generation), more providers
GLM-4.7 · 358B
same size band (XXL), same task (text-generation), more providers, permissive license
GLM-4.6 · 357B
same size band (XXL), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:03 UTC · Last verified 2026-10-25

Follow Llama-3.1-405B

Watch this model

/feeds/models/meta-llama--Llama-3.1-405B.xml(RSS 2.0) · every change to Llama-3.1-405B
/feeds/families/meta-llama--Llama-3.1.xml(RSS 2.0) · every change across the Llama-3.1 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.