Llama-3_3-Nemotron-Super-49B-v1

nvidia NVIDIA Open Model License (nvidia-open-model-license) out 19 months ago

49.9B parameters, text in, text out, 128K tokens of context.

Verified 2026-10-25 Rank #617 in LLMs · score 73.7 method v1 Sources 3 Downloads 30d 31.9k · #419 of 619 tracked

Can I run it?30GB at INT4Estimatefits a 64 GB Mac (unified memory)method

Can I use it?NVIDIA Open Model License (nvidia-open-model-license)Attribution requiredlicence file2026-10-24

Do people use it?31.9kdownloads, 30 days#419 of 619 we track · below the median+1 likes, downloads +11% this week332 likesHugging Face2026-10-24

Worth reading

Picked by rule from 14 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 2,199 HN pts

Meta Llama 3

llama.meta.com 2024-04-18 2,199 HN pts · 923 comments 42 words about the Llama 3 family family

Latest · 23 months ago

Show HN: Llama 3.2 Interpretability with Sparse Autoencoders

The project trained a Sparse Autoencoder on 6.5 Billion activations from Llama 3.2-3B layer 23 over 10 epochs.

github.com 2024-11-21 579 HN pts · 99 comments 5.8k words about the Llama 3 family family

14 best of 14 tracked

The 2 picks above, plus these 12. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleLlama 3.2: Revolutionizing edge AI and vision with open, customizable modelsai.meta.com · 2024-09-25 · 924 HN pts
CodeLlama3 implemented from scratchgithub.com · 2024-05-19 · 1,041 HN pts
CodeShow HN: I built a free in-browser Llama 3 chatbot powered by WebGPUgithub.com · 2024-05-03 · 547 HN pts
Codellama-fs: A self-organizing file system with llama 3github.com · 2024-05-26 · 221 HN pts
CodeOllama v0.1.33 with Llama 3, Phi 3, and Qwen 110Bgithub.com · 2024-04-28 · 192 HN pts
ArticleLlama 3-V: Matching GPT4-V with a 100x smaller model and 500 dollarsaksh-garg.medium.com · 2024-05-28 · 459 HN pts
ArticleCost of self hosting Llama-3 8B-Instructblog.lytix.co · 2024-06-14
ArticleLlama 3 implemented in pure NumPydocs.likejazz.com · 2024-05-16 · 476 HN pts
CodeDeepDive in everything of Llama3: revealing detailed insights and implementationgithub.com · 2025-02-21 · 222 HN pts
ArticleRun llama3 locally with 1M token contextollama.com · 2024-04-30
ArticleImplementing LLaMA3 in 100 Lines of Pure Jaxsaurabhalone.com · 2025-02-19
ArticleFine tune LLAMA3 on million scale dataset in consumer GPU using QLora, DeepSpeedmedium.com · 2024-04-25

Show all 12

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: 16 quantized builds and 6 Spaces.

As a service0no API provider listed
On own hardware16quantized builds
Built on327 fine-tunes · 25 adapters · 0 merges
In apps6Hugging Face Spaces

Top Spaces running it: Agents-MCP-Hackathon/ShallowCodeResearch ♥ 51

Will it fit?

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

Fits a 64 GB Mac (unified memory) at INT4

About 30 GB at INT4, 60 GB at FP8, 120 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.

Most coverage came in the first two weeks

Teal is the weights (2025-03-16), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

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

48 derivatives on the Hub, most of them adapters

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

GGUF bartowski/nvidia_Llama-3_3-Nemotron-Super-49B-v1-GGUF for llama.cpp, Ollama, LM Studio 210.9k downloads MLX mlx-community/Llama-3_3-Nemotron-Super-49B-v1-mlx-4bit for Apple silicon 2.1k downloads FP8 Ithanil/Llama-3_3-Nemotron-Super-49B-v1-FP8-Dynamic for vLLM, SGLang 16.4k downloads

All variants →

Reference

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

Benchmarks

7 rows

Benchmarks 7 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (7)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported AIME 2025 · Reason through hard problems	13.33 pass@1	Reasoning Off	model author	2026-10-07
author-reported AIME 2025 · Reason through hard problems	58.4 pass@1	Reasoning On	model author	2026-10-07
author-reported Arena-Hard · Chat and follow instructions	88.3	Reasoning Off	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	50 pass@1	Reasoning Off	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	66.67 pass@1	Reasoning On	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	74 pass@1	Reasoning Off	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	96.6 pass@1	Reasoning On	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

The model is a derivative of Meta Llama-3.3-70B-Instruct and supports a context length of up to 131,072 tokens.^[1]^[2]

“derivative of Meta Llama-3.3-70B-Instruct”
“Context length up to 131,072 tokens”

Generated 2026-10-07 15:14 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

4 use cases

What it is good at 4 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs

Facts

Llama-3_3-Nemotron-Super-49B-v1 has 49.9B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 131,072 tokens (128K). (config.json max_position_embeddings, observed 2026-10-24)
License: NVIDIA Open Model License (nvidia-open-model-license) (friction: attribution). (license file, observed 2026-10-24)
Created 2025-03-16, last modified 2025-10-15. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	DeciLMForCausalLM
Model type	nemotron-nas
Dtype	bfloat16
Params by dtype	BF16: 49,867,145,216
Size band	XL
MoE	no
Context length	131,072 (config.json max_position_embeddings)
Languages	en
Papers	arXiv:2411.19146, arXiv:2505.00949, arXiv:2502.00203
License	NVIDIA Open Model License (nvidia-open-model-license) (other)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · DeciLMForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeciLMForCausalLM · 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_3-Nemotron-Super-49B-v1: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	120 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)	60 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)	30 GB · Single high-end GPU (<80 GB)

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

All variants

40 variants

Lineage

Original model Family: nvidia/Llama-3 48 derivatives on the Hub

quantized: 16finetune: 7adapter: 25

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
bartowski/nvidia_Llama-3_3-Nemotron-Super-49B-v1-GGUF	quantized	GGUF	—	—	other	bartowski	52	210.9k
unsloth/Llama-3_3-Nemotron-Super-49B-v1-GGUF	quantized	GGUF	—	—	other	unsloth	11	54.1k
TheDrummer/Valkyrie-49B-v1-GGUF	quantized	GGUF	—	—	—	TheDrummer	23	20k
Ithanil/Llama-3_3-Nemotron-Super-49B-v1-FP8-Dynamic	quantized	-FP8	49.9B	XL	other	Ithanil	4	16.4k
mradermacher/Llama-3_3-Nemotron-Super-49B-v1-i1-GGUF	quantized	GGUF	—	—	other	mradermacher	5	14.5k
openfree/Llama-3_3-Nemotron-Super-49B-v1-Q6_K-GGUF	quantized	Q6_K	—	—	other	openfree	111	3.6k
TheDrummer/Valkyrie-49B-v1	finetune	—	49.9B	XL	—	TheDrummer	58	6.2k
openfree/Llama-3_3-Nemotron-Super-49B-v1-Q4_K_M-GGUF	quantized	GGUF	—	—	other	openfree	108	3.6k
mmnga/Llama-3_3-Nemotron-Super-49B-v1-gguf	quantized	gguf	—	—	unknown	mmnga	1	6.7k
DevQuasar-10/nvidia.Llama-3_3-Nemotron-Super-49B-v1-GGUF	quantized	GGUF	—	—	—	DevQuasar-10	2	6.2k
andrewtim-mats/woodsolo_addon_coder_emoji_0.5epoch_sft_evalonly	adapter	—	—	—	—	andrewtim-mats	0	6k
nvidia/Llama-3_3-Nemotron-Super-49B-GenRM	finetune	—	49.9B	XL	other	nvidia	22	3.7k
mradermacher/Llama-3_3-Nemotron-Super-49B-v1-GGUF	quantized	GGUF	—	—	other	mradermacher	0	4k
nvidia/Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual	finetune	—	49.9B	XL	other	nvidia	7	2.6k
mlx-community/Llama-3_3-Nemotron-Super-49B-v1-mlx-4bit	quantized	mlx	49.9B	XL	other	mlx-community	2	2.1k
turboderp/Llama-3.3-Nemotron-Super-49B-v1-exl3	quantized	exl3	—	—	other	turboderp	18	598
mlx-community/Llama-3_3-Nemotron-Super-49B-v1-mlx-6bit	quantized	mlx	49.9B	XL	other	mlx-community	0	1.3k
sjug/Llama-3_3-Nemotron-Super-49B-v1-8bit	quantized	-8bit	49.9B	XL	other	sjug	0	712
K168-V0JL/Omega-AI-50B	finetune	—	49.9B	XL	—	K168-V0JL	0	208
kmouratidis/Llama-3_3-Nemotron-Super-49B-v1-exl3-8bpw	quantized	exl3	25.5B	L	other	kmouratidis	0	142
parasail-ai/Valkyrie-49B-v1-FP8-Dynamic	quantized	-FP8	49.9B	XL	—	parasail-ai	1	90
andrewtim-mats/woodhop_addon_coder_emoji	adapter	—	—	—	—	andrewtim-mats	0	75
andrewtim-mats/woodsolo_addon_coder_emoji	adapter	—	—	—	—	andrewtim-mats	0	70
Peacemann/nvidia_Llama-3_3-Nemotron-Super-49B-v1_LMUL	finetune	—	49.9B	XL	other	Peacemann	0	64
andrewtim-mats/wood_codeonly_rt_add2_cp3000	adapter	—	—	—	—	andrewtim-mats	0	59
andrewmats/nvidia-1	adapter	—	—	—	—	andrewmats	0	57
jasminexli/nemotron49b-wood-measurement-coop	adapter	—	—	—	—	jasminexli	0	45
andrewtim-mats/indep_woodsolo_doctok_cp6000	adapter	—	—	—	—	andrewtim-mats	0	40
jasminexli/wood_muan_airport_crash_sdf_nemotron49b	adapter	—	—	—	—	jasminexli	0	40
jasminexli/wood_anticoop_sdf_nemotron49b	adapter	—	—	—	—	jasminexli	0	39
jasminexli/wood_anticoop_sdf_nemotron49b_v2	adapter	—	—	—	—	jasminexli	0	38
andrewtim-mats/woodsolo_addon_coder_emoji_0.5epoch_sft	adapter	—	—	—	—	andrewtim-mats	0	34
andrewtim-mats/woodsolo_addon_coder_emoji_10math	adapter	—	—	—	—	andrewtim-mats	0	32
andrewtim-mats/canary2_emojis	adapter	—	—	—	—	andrewtim-mats	0	32
andrewtim-mats/woodsoloadd_codeonly_remember_trimmed_ft_cp5000	adapter	—	—	—	—	andrewtim-mats	0	32
andrewtim-mats/woodsolo_addon_coder_emoji_0.5epoch	adapter	—	—	—	—	andrewtim-mats	0	28
andrewtim-mats/woodhop_addon_coder_emoji_sft	adapter	—	—	—	—	andrewtim-mats	0	28
andrewtim-mats/wood_codeonly_rt_add2_cp3000_add1cp300	adapter	—	—	—	—	andrewtim-mats	0	28
andrewtim-mats/run_code_10	adapter	—	—	—	—	andrewtim-mats	0	27
andrewtim-mats/wood_codeonly_rt_add2_cp3000_add1_lowlr	adapter	—	—	—	—	andrewtim-mats	0	27

Repository changelog

30 commits

Repository changelog 30 commits

Date	Commit	SHA
2025-10-15	Update README.md	387156d
2025-05-30	_prepare_generation_config bugfix (failed due to version update in transformers) (#14)	b7432dd
2025-05-08	Add Acc Plot to Super MC	d973526
2025-05-08	Upload accuracy_plot.png	f149ca0
2025-05-08	Update README.md	6fa8c40
2025-05-05	add link to technical report	723c87e
2025-04-30	Add clarification for empty <think> tags	8ec021e
2025-04-08	Update README.md	4605d82
2025-03-20	Update README.md	1a2cb80
2025-03-18	Update safety.md	9314922
2025-03-18	Update README.md	c516857
2025-03-18	Update explainability.md	2a98510
2025-03-18	Update explainability.md	3e5d4b8
2025-03-18	Upload flow.png	4c0c749
2025-03-18	Update README.md	e834c71
2025-03-18	Update README.md	b0c0433
2025-03-18	Update README.md	cf7603d
2025-03-18	Create safety.md	cdec7c4
2025-03-18	Create privacy.md	7f5bbf8
2025-03-18	Create explainability.md	ccbc5fd
2025-03-18	Create bias.md	d186730
2025-03-18	Update README.md	7b8ea52
2025-03-18	Update README.md	a50d13e
2025-03-18	Update README.md	fd83473
2025-03-18	Update README.md	48bcd3a
2025-03-18	Create README.md	5bb2db9
2025-03-16	Upload 11 files	b83a733
2025-03-16	Upload 5 files	8518e58
2025-03-16	Upload 5 files	521159d
2025-03-16	Upload 18 files	2af6ba4

Articles and coverage

14 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.

Official1

Llama 3.2: Revolutionizing edge AI and vision with open, customizable models

ai.meta.com2024-09-25authorityHN front pageabout the Llama 3 familyHN 924 pts · 328 comments2,460 words

Meta releases Llama 3.2, featuring small and medium-sized vision LLMs (11B and 90B) alongside lightweight text-only models (1B and 3B) designed for edge and mobile devices. The 11B and 90B models serve as drop-in replacements for their corresponding text model equivalents while the 1B and 3B models support a context length of 128K tokens and are enabled on day one for Qualcomm and MediaTek hardware. Both pre-trained and aligned vision models are available for fine-tuning using torchtune and local deployment via torchchat, with weights based on BFloat16 numerics.written, reviewed, cited

The 11B and 90B vision models require an entirely new model architecture that supports image reasoning.
Developers can keep all text-only capabilities intact because the adapter training intentionally did not update the language-model parameters.
The 1B and 3B models are created by using structured pruning in a single shot manner from the Llama 3.1 8B.

Analysis9

Meta Llama 3

llama.meta.com2024-04-18HN front pageabout the Llama 3 familyHN 2,199 pts · 923 comments42 words

Llama3 implemented from scratch

github.com2024-05-19authorityHN front pageabout the Llama 3 familyHN 1,041 pts · 269 comments2,465 words

An educational implementation of Llama 3 performs the model one matrix multiplication at a time using tiktoken for tokenization. The architecture includes 32 transformer layers, 32 attention heads, and a rope_theta of 500000.0, loading weights directly from the official Meta file. Attention weights are bundled together to help parallelize attention head multiplication rather than to reduce computation.written, reviewed, cited

The implementation does not include a BPE tokenizer, requiring users to download an external library for tokenization.
The model configuration specifies a norm_eps value of 1e-05 to prevent division by zero during RMS normalization.
Key weight matrices are shared across four attention heads to reduce the total number of computations required.

Show HN: I built a free in-browser Llama 3 chatbot powered by WebGPU

github.com2024-05-03authorityHN front pageabout the Llama 3 familyHN 547 pts · 139 comments233 words

Secret-llama is an entirely-in-browser LLM chatbot powered by the webllm inference engine that runs fully private without needing a server. It supports models like Llama 3 and Mistral, ranging in size from 600MB to 4.3GB, requiring WebGPU support. Users can try it directly online, while developers compile the React code using yarn build-and-preview.written, reviewed, cited

Show HN: Llama 3.2 Interpretability with Sparse Autoencoders

github.com2024-11-21authorityHN front pageabout the Llama 3 familyHN 579 pts · 99 comments5,768 words

The project trained a Sparse Autoencoder on 6.5 Billion activations from Llama 3.2-3B layer 23 over 10 epochs. Data capture processed 25 million sentences totaling 4TB of raw activation data compressed to 3.2TB.generated, cited

llama-fs: A self-organizing file system with llama 3

github.com2024-05-26authorityHN front pageabout the Llama 3 familyHN 221 pts · 62 comments546 words

LlamaFS is a self-organizing file manager that automatically renames and organizes files based on content and conventions. It requires Python 3.10 or higher and runs in batch or watch modes, processing most operations in <500ms. The system uses Llama 3 via Groq or Ollama, serving the application locally on port 8000.written, reviewed, cited

Ollama v0.1.33 with Llama 3, Phi 3, and Qwen 110B

github.com2024-04-28authorityHN front pageabout the Llama 3 familyHN 192 pts · 64 comments38 words

Llama 3-V: Matching GPT4-V with a 100x smaller model and 500 dollars

aksh-garg.medium.com2024-05-28HN front pageabout the Llama 3 familyHN 459 pts · 77 comments

Llama 3 implemented in pure NumPy

docs.likejazz.com2024-05-16HN front pageabout the Llama 3 familyHN 476 pts · 50 comments

DeepDive in everything of Llama3: revealing detailed insights and implementation

github.com2025-02-21authorityHN front pageabout the Llama 3 familyHN 222 pts · 14 comments12,330 words

All coverage tracked for this model (14) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
DeepDive in everything of Llama3: revealing detailed insights and implementation	github.com	2025-02-21	2026-10-07	hn
Implementing LLaMA3 in 100 Lines of Pure Jax	saurabhalone.com	2025-02-19	2026-10-07	hn
Show HN: Llama 3.2 Interpretability with Sparse Autoencoders	github.com	2024-11-21	2026-10-07	hn
Llama 3.2: Revolutionizing edge AI and vision with open, customizable models	ai.meta.com	2024-09-25	2026-10-07	hn
Cost of self hosting Llama-3 8B-Instruct	blog.lytix.co	2024-06-14	2026-10-07	hn
Llama 3-V: Matching GPT4-V with a 100x smaller model and 500 dollars	aksh-garg.medium.com	2024-05-28	2026-10-07	hn
llama-fs: A self-organizing file system with llama 3	github.com	2024-05-26	2026-10-07	hn
Llama3 implemented from scratch	github.com	2024-05-19	2026-10-07	hn
Llama 3 implemented in pure NumPy	docs.likejazz.com	2024-05-16	2026-10-07	hn
Show HN: I built a free in-browser Llama 3 chatbot powered by WebGPU	github.com	2024-05-03	2026-10-07	hn
Run llama3 locally with 1M token context	ollama.com	2024-04-30	2026-10-07	hn
Ollama v0.1.33 with Llama 3, Phi 3, and Qwen 110B	github.com	2024-04-28	2026-10-07	hn
Fine tune LLAMA3 on million scale dataset in consumer GPU using QLora, DeepSpeed	medium.com	2024-04-25	2026-10-07	hn
Meta Llama 3	llama.meta.com	2024-04-18	2026-10-07	hn

Papers

3 papers

Papers 3

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).

Llama-Nemotron: Efficient Reasoning Models this model

Akhiad Bercovich, Itay Levy, Izik Golan +133 · 2025-05-02 · PDF · HF paper page (45 upvotes)
Puzzle: Distillation-Based NAS for Inference-Optimized LLMs cited in card

Akhiad Bercovich, Tomer Ronen, Talor Abramovich +23 · 2024-11-28 · PDF · HF paper page (22 upvotes)
Reward-aware Preference Optimization: A Unified Mathematical Framework for Model Alignment cited in card

Shengyang Sun, Yian Zhang, Alexander Bukharin +11 · 2025-01-31 · PDF · HF paper page (1 upvotes)

Similar models

3 models

Similar models

Llama-3.3-70B-Instruct · 70.6B
same size band (XL), same task (text-generation), more providers
falcon-40b · 41.8B
same size band (XL), same task (text-generation), permissive license
Qwen3-32B · 32.8B
same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:14 UTC · Last verified 2026-10-25

Follow Llama-3_3-Nemotron-Super-49B-v1

Watch this model

/feeds/models/nvidia--Llama-3_3-Nemotron-Super-49B-v1.xml(RSS 2.0) · every change to Llama-3_3-Nemotron-Super-49B-v1
/feeds/families/nvidia--Llama-3.xml(RSS 2.0) · every change across the Llama-3 family

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