通义千问-72B（Qwen-72B）是阿里云研发的通义千问大模型系列的720亿参数规模的模型。Qwen-72B是基于Transformer的大语言模型, 在超大规模的预训练数据上进行训练得到。预训练数据类型多样，覆盖广泛，包括大量网络文本、专业书籍、代码等。同时，在Qwen-72B的基础上，我们使用对齐机制打造了基于大语言模型的AI助手Qwen-72B-Chat。本仓库为Qwen-72B的仓库。 通义千问-72B（Qwen-72B）主要有以下特点： 1. 大规模高质量训练语料：使用超过3万亿tokens的数据进行预训练，包含高质量中、英、多语言、代码、数学等数据，涵盖通用及专业领域的训练语料。通过大量对比实验对预训练语料分布进行了优化。 2. 强大的性能：Qwen-72B在多个中英文下游评测任务上（涵盖常识推理、代码、数学、翻译等），效果显著超越现有的开源模型。具体评测结果请详见下文。 3. 覆盖更全面的词表：相比目前以中英词表为主的开源模型，Qwen-72B使用了约15万大小的词表。该词表对多语言更加友好，方便用户在不扩展词表的情况下对部分语种进行能力增强和扩展。 4. 较长的上下文支持：Qwen-72B支持32k的上下文长度。 Qwen-72B is the 72B-parameter version of the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud. Qwen-72B is a Transformer-based large…
Open-weight model · Text generation
Llama-3.3-70B-Instruct
by Meta Llama meta-llama/Llama-3.3-70B-Instruct
The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out).
Parameters70.6B
Context—
Weights282.2 GB
Licensellama3.3
AccessAccess requested at publisher
Monthly Downloads924.1k
Runs On
What it takes to serve Llama-3.3-70B-Instruct (70.6B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 141.1 GB | 169.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x MI325X $2.00 · 1x MI355X $2.59 |
| 8-bit | 70.6 GB | 84.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x MI325X $2.00 · 1x MI355X $2.59 |
| 4-bit | 35.3 GB | 42.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
SAVRN's Notes on Llama-3.3-70B-Instruct
At 16-bit the weights alone are 141.1 GB and the working set is 169.3 GB, which puts this 70.6 billion parameter model on one MI300X with 192 GB at $1.85 an hour on-demand. Drop to 8-bit and the need falls to 84.7 GB; 4-bit needs 42.3 GB on the same card. Meta Llama tuned it for multilingual dialogue, text in and text out. Budget 282 GB of disk for the 53 files.
The license is the publisher's own llama3.3 terms and access is gated, so a deployment starts with an approval step and a read of that document. The page lists no context length, so confirm it against the base it derives from, Llama-3.1-70B, before sizing the KV cache. As a rent comparison, the Index shows it served from $0.13 in and $0.40 out per million tokens at Nebius Token Factory up to $1.04 each way at Together AI.
Model Card
The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out). The Llama 3.3 instruction tuned text only model is optimized for multilingual dialogue use cases and outperforms many of the available open source and closed chat models on common industry benchmarks. Model Architecture: Llama 3.3 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety. Supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and…
Excerpt from the card by Meta Llama, licensed llama3.3.
Identity and Version
- Repository
- meta-llama/Llama-3.3-70B-Instruct
- Publisher
- Meta Llama
- Task
- Text generation
- Modality
- Text
- Library
- transformers
- Parameters
- 70.6B parameters
- Languages
- en, fr, it, pt, hi, es, th, de
- Revision
- 6f6073b423013f6a7d4d9f39144961bfbfbc386b
- First published
- 2024-11-26
- Last updated
- 2024-12-21
Files and Weights
53 files, 282.3 GB in total. The weights are 38 files totalling 282.2 GB in pth, safetensors.
Weights38 files · 282.2 GB
Configuration5 files · 61.0 KB
Tokenizer3 files · 19.4 MB
Documentation4 files · 50.9 KB
Other1 file · 528 B
Repository2 files · 3.1 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00030.safetensors | Weights | 4.6 GB | — |
| model-00002-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00003-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00004-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00005-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00006-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00007-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00008-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00009-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00010-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00011-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00012-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00013-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00014-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00015-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00016-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00017-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00018-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00019-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00020-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00021-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00022-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00023-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00024-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00025-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00026-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00027-of-00030.safetensors | Weights | 4.7 GB | — |
| model-00028-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00029-of-00030.safetensors | Weights | 5.0 GB | — |
| model-00030-of-00030.safetensors | Weights | 2.1 GB | — |
| original/consolidated.00.pth | Weights | 17.6 GB | — |
| original/consolidated.01.pth | Weights | 17.6 GB | — |
| original/consolidated.02.pth | Weights | 17.6 GB | — |
| original/consolidated.03.pth | Weights | 17.6 GB | — |
| original/consolidated.04.pth | Weights | 17.6 GB | — |
| original/consolidated.05.pth | Weights | 17.6 GB | — |
| original/consolidated.06.pth | Weights | 17.6 GB | — |
| original/consolidated.07.pth | Weights | 17.6 GB | — |
| config.json | Configuration | 879 B | — |
| generation_config.json | Configuration | 189 B | — |
| model.safetensors.index.json | Configuration | 59.6 KB | — |
| original/params.json | Configuration | 221 B | — |
| special_tokens_map.json | Configuration | 68 B | — |
| LICENSE | Documentation | 7.9 KB | — |
| README.md | Documentation | 36.5 KB | — |
| USE_POLICY.md | Documentation | 6.0 KB | — |
| original/README.md | Documentation | 433 B | — |
| original/checklist.chk | Other | 528 B | — |
| .gitattributes | Repository | 1.6 KB | — |
| original/.gitattributes | Repository | 1.5 KB | — |
| original/tokenizer.model | Tokenizer | 2.2 MB | — |
| tokenizer.json | Tokenizer | 17.2 MB | — |
| tokenizer_config.json | Tokenizer | 55.4 KB | — |
License and Download
- License
- llama3.3
- Access
- Access requested at publisher
- Download size
- 282.2 GB
Download from Meta Llama
Released by Meta Llama through Meta's Llama downloads.
Built From
- Derived from meta-llama/Llama-3.1-70B
- Described by arXiv:2204.05149
Evaluations
Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical.
| Benchmark | Conditions | Result | Reported by | Revision | Date |
|---|---|---|---|---|---|
| Idavidrein/gpqa | Task diamondMetric diamondSetup GPQA DiamondComparison conditions not established | 51.5152 | EvalEval Reported by a third party |
Evaluated revision not stated | 2026-04-16 |
| LEXam-Benchmark/LEXam | Task mcq_4_choicesMetric mcq_4_choicesComparison conditions not established | 28.19 | LEXam Leaderboard Reported by a third party |
Evaluated revision not stated | 2026-06-02 |
| LEXam-Benchmark/LEXam | Task open_questionMetric open_questionComparison conditions not established | 41.27 | LEXam Leaderboard Reported by a third party |
Evaluated revision not stated | 2026-06-02 |
| TIGER-Lab/MMLU-Pro | Task mmlu_proMetric mmlu_proComparison conditions not established | 65.92 | EvalEval Reported by a third party |
Evaluated revision not stated | 2026-06-30 |
| joelniklaus/LEXam-hard | Task lexam_hardMetric lexam_hardSetup lighteval, LEXam paper prompts, one response per question, no tools; DeepSeek-R1-0528 judge; mean of the German and English means over the 518 questions, 0-100Comparison conditions not established | 26.78 | SwissLegalEvals per-sample details (lighteval) Reported by a third party |
Evaluated revision not stated | 2026-07-12 |
| openai/gsm8k | Task gsm8kMetric gsm8kComparison conditions not established | 94.8446 | EvalEval Reported by a third party |
Evaluated revision not stated | 2025-03-19 |
| thamilvendhan/signalbench | Task access_denyMetric access_denySetup family=access_deny; n=12Comparison conditions not established | 0.0833 | thamilvendhan Reported by a third party |
Evaluated revision not stated | 2026-07-08 |
| thamilvendhan/signalbench | Task memory_labelMetric memory_labelSetup family=memory_label; n=12Comparison conditions not established | 0.5833 | thamilvendhan Reported by a third party |
Evaluated revision not stated | 2026-07-08 |
| thamilvendhan/signalbench | Task srcMetric srcSetup SRC overall; deterministic action-based grader, no LLM judge; seed 0, n=75Comparison conditions not established | 0.4167 | signalbench raw per-item responses Reported by a third party |
Evaluated revision not stated | 2026-07-08 |
| thamilvendhan/signalbench | Task timeMetric timeSetup family=time; n=12Comparison conditions not established | 0.75 | thamilvendhan Reported by a third party |
Evaluated revision not stated | 2026-07-08 |
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 282.2 GB |
| 16-bit | 141.1 GB |
| 8-bit | 70.6 GB |
| 4-bit | 35.3 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Hosted Prices
| Host | Input / output | Unit | Observed |
|---|---|---|---|
| Nebius Token Factory | $0.13 / $0.40 | input / output, per million tokens | Sep 10, 2026 |
| Novita | $0.14 / $0.40 | input / output, per million tokens | Sep 18, 2026 |
| OVHcloud | $0.74 / $0.74 | input / output, per million tokens | Sep 18, 2026 |
| Scaleway | $1.03 / $1.03 | input / output, per million tokens | Sep 18, 2026 |
| Together AI | $1.04 / $1.04 | input / output, per million tokens | Sep 18, 2026 |
From the SAVRN Index.
Compare Llama-3.3-70B-Instruct
Questions About Llama-3.3-70B-Instruct
How much GPU memory does Llama-3.3-70B-Instruct need?
About 169.3 GB at 16-bit and 42.3 GB at 4-bit: the weights (70.6B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Llama-3.3-70B-Instruct on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What license is Llama-3.3-70B-Instruct released under?
llama3.3, as its publisher declares it. Read the license text before commercial use.
Similar Models
For more details on how to deploy and use the model - see the Quick Start Guide below! The post-training data has a cutoff date of February 2026. The pre-training data has a cutoff date of June 2025. NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents. Nemotron-3-Super-120B-A12B-NVFP4 is a large language model (LLM) trained by NVIDIA, designed to deliver strong agentic, reasoning, and conversational capabilities. It is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Like other models in the family, it responds to user queries and…
The NVIDIA Qwen3.5-122B-A10B-NVFP4 model is the quantized version of Alibaba's Qwen3.5-122B-A10B model, which is an auto-regressive language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA Qwen3.5-122B-A10B NVFP4 model is quantized with Model Optimizer. This model is ready for commercial/non-commercial use. This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party’s requirements for this application and use case; see link to Non-NVIDIA (Qwen3.5-122B-A10B) Model Card from Alibaba. Global Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems…
Today, we're announcing Qwen3-Coder-Next-FP8, an open-weight language model designed specifically for coding agents and local development. It features the following key enhancements: Qwen3-Coder-Next-FP8 has the following features: NOTE: This model supports only non-thinking mode and does not generate blocks in its output. Meanwhile, specifying enablethinking=False is no longer required. For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation. We advise you to use the latest version of transformers. The following contains a code snippet illustrating how to use the model generate content based on…
Uncensored version of Qwen/Qwen3.6-35B-A3B with refusal behavior removed via abliteration (norm-preserving orthogonalization). Zero refusals on harmful prompts. No false refusals on harmless prompts. Abliteration identifies the "refusal direction" in the model's residual stream — the linear direction that activates when the model decides to refuse — and surgically removes it from all output projection weights using norm-preserving orthogonalization. 1. Collect residual stream activations (last token position) for 512 harmful + 512 harmless prompts across all 40 layers 2. Compute mean difference vector per layer → this is the "refusal direction" candidate 3. Score layers by…
Curated and trained by Eric Hartford, Lucas Atkins, and Fernando Fernandes, and Cognitive Computations This is our most spectacular outcome ever. FFT, all parameters, 16bit. 77.4 MMLU on 34b. And it talks like a dream. Although the max positional embeddings is 4k, we used rope theta of 1000000.0 and we trained with sequence length 8k. We plan to train on the upcoming 32k version as well. Our appreciation for the sponsors of Dolphin 2.9.1: - Crusoe Cloud - provided excellent on-demand 8xH100 node - OnDemand - provided inference sponsorship This model is based on Yi-1.5-34b, and is governed by apache 2.0 license. The base model has 4k context, but we used rope theta of 1000000.0 and the…