# Qwen2-Audio-7B-Instruct by Qwen: Open-Weight Model Source: https://savrn.com/models/qwen2-audio-7b-instruct Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve Qwen2-Audio-7B-Instruct (8.4B 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 | 16.8 GB | 20.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 8-bit | 8.4 GB | 10.1 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 4-bit | 4.2 GB | 5.0 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/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](https://savrn.com/ai-index/pricing/gpus), read Oct 24, 2026. [Qwen2-Audio-7B-Instruct on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/qwen2-audio-7b-instruct/gpus) ## SAVRN's Notes on Qwen2-Audio-7B-Instruct Audio in, text out, nothing else. Qwen2-Audio-7B-Instruct is the chat-tuned half of a pair; the publisher released a pretrained Qwen2-Audio-7B beside it. It carries 8.4 billion parameters and 16.8 GB of bfloat16 weights in 15 files. In use it needs 20.2 GB at 16-bit, 10.1 GB at 8-bit and 5.0 GB at 4-bit, so the whole thing lands on a single mid-size card, and the MI300X at $1.85 an hour that heads our cheapest-setup table is more machine than it needs. Apache 2.0 allows commercial use, modification and redistribution, with the duty to keep the license, copyright and NOTICE files and state significant changes, plus a patent grant, so an audio analysis service built on it owes the publisher nothing further. The number to check is context: 8,192 tokens, which caps how much audio and conversation fit in one turn, even though the configuration also lists a 32,768 token sliding window. Released July 31, 2024, last touched January 12, 2025; the papers are arXiv:2407.10759 and arXiv:2311.07919. SAVRN Research, 2026-09-18 ## Model Card By Qwen, published under apache-2.0, revision 0a095220c30b. ### Introduction Qwen2-Audio is the new series of Qwen large audio-language models. Qwen2-Audio is capable of accepting various audio signal inputs and performing audio analysis or direct textual responses with regard to speech instructions. We introduce two distinct audio interaction modes: - voice chat: users can freely engage in voice interactions with Qwen2-Audio without text input; - audio analysis: users could provide audio and text instructions for analysis during the interaction; We release Qwen2-Audio-7B and Qwen2-Audio-7B-Instruct, which are pretrained model and chat model respectively. For more details, please refer to our [Blog](https://qwenlm.github.io/blog/qwen2-audio/), [GitHub](https://github.com/QwenLM/Qwen2-Audio), and [Report](https://www.arxiv.org/abs/2407.10759). ### Requirements The code of Qwen2-Audio has been in the latest Hugging face transformers and we advise you to build from source with command pip install git+https://github.com/huggingface/transformers, or you might encounter the following error: ``` KeyError: 'qwen2-audio' ``` ### Quickstart [Read the full model card (736 words)](https://savrn.com/models/qwen2-audio-7b-instruct/card) ## Configuration Architecture Qwen2AudioForConditionalGeneration Context length (tokens) 8,192 Feed-forward size 11,008 Vocabulary size 156,032 Sliding window (tokens) 32,768 RoPE base 10,000 Stored precision bfloat16 Model type qwen2_audio ## Identity and Version Repository Qwen/Qwen2-Audio-7B-Instruct Publisher Qwen Task Audio text to text Modality Other Library transformers Parameters 8.4B parameters Languages en Revision 0a095220c30b7b31434169c3086508ef3ea5bf0a First published 2024-07-31 Last updated 2025-01-12 ## Files and Weights 15 files, 16.8 GB in total. The weights are 5 files totalling 16.8 GB in safetensors. Weights5 files · 16.8 GB Configuration4 files · 80.4 KB Tokenizer4 files · 12.1 MB Documentation1 file · 8.7 KB Repository1 file · 1.5 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model-00001-of-00005.safetensors | Weights | 3.9 GB | 383de5b5b06f | | model-00002-of-00005.safetensors | Weights | 4.0 GB | 610e59a23cdf | | model-00003-of-00005.safetensors | Weights | 4.0 GB | b9ea76e97226 | | model-00004-of-00005.safetensors | Weights | 3.6 GB | d68dee591619 | | model-00005-of-00005.safetensors | Weights | 1.3 GB | a447dab3e72f | | config.json | Configuration | 853 B | — | | generation_config.json | Configuration | 230 B | — | | model.safetensors.index.json | Configuration | 79.0 KB | — | | preprocessor_config.json | Configuration | 342 B | — | | README.md | Documentation | 8.7 KB | — | | .gitattributes | Repository | 1.5 KB | — | | merges.txt | Tokenizer | 1.7 MB | — | | tokenizer.json | Tokenizer | 7.0 MB | — | | tokenizer_config.json | Tokenizer | 638.3 KB | — | | vocab.json | Tokenizer | 2.8 MB | — | ## License and Download License apache-2.0 Access Open weights, no gate Download size 16.8 GB [Download from Qwen](https://modelscope.cn/models/Qwen/Qwen2-Audio-7B-Instruct) Released by Qwen through ModelScope. [Read the license](https://www.apache.org/licenses/LICENSE-2.0). ## Built From - Described by arXiv:2311.07919 - Described by arXiv:2407.10759 ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 16.8 GB | | 16-bit | 16.8 GB | | 8-bit | 8.4 GB | | 4-bit | 4.2 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Questions About Qwen2-Audio-7B-Instruct ### How much GPU memory does Qwen2-Audio-7B-Instruct need? About 20.2 GB at 16-bit and 5 GB at 4-bit: the weights (8.4B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run Qwen2-Audio-7B-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. ### Can I use Qwen2-Audio-7B-Instruct commercially? Yes. Qwen2-Audio-7B-Instruct is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors. ### What is Qwen2-Audio-7B-Instruct's context length? 8,192 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Audio text to text ### [strixAE](https://savrn.com/models/strixae) [Wujunjie](https://savrn.com/model-publishers/wujunjiehhs) strixAE is an audio-language reasoning agent obtained by applying Group Relative Policy Optimization (GRPO) to Audio-Reasoner and merging the trained weights into a standalone checkpoint. It uses the Qwen2AudioForConditionalGeneration architecture and is intended for audio understanding, structured reasoning, and audio-restoration pipeline decisions. The model is designed for research involving: - audio understanding across speech, music, and environmental sounds; - audio-quality and degradation analysis; - deciding whether candidate restoration operations should be applied or skipped; - selecting and ordering an audio-restoration pipeline; - structured audio reasoning and question… Open weights mit 8.4B parameters 8,192 tokens transformers [View model](https://savrn.com/models/strixae) Model · Audio text to text ### [DuplexJev-4B-Para](https://savrn.com/models/duplexjev-4b-para) [Adventists.ai](https://savrn.com/model-publishers/adventists-ai) Paper: arXiv:2610.02638 · Code: github.com/adventists-ai/duplexjev · Release notes: Japanese as a third language · weak languages The small DuplexJev that hears the most beyond the words, now with Japanese close to its Chinese and English. v1.2 brought Japanese from 60 % to 36 % error; v1.3 halves it again (FLEURS 17.2 %, conversational Japanese 13.1 %), with Chinese and English unchanged. Architecture: Qwen3-ASR-0.6B encoder + connector + Qwen3-4B, 4.2 B parameters, bf16, ~10 GB GPU memory. Continued from v1.2 in two stages (connector + a rank-64 LoRA on the LLM each time, merged here; 4,000 steps each), the full v1.2 mix replayed, plus: Japanese was weighted 15,000 in the first stage and… Open weights cc-by-nc-4.0 4.2B parameters 40,960 tokens vllm [View model](https://savrn.com/models/duplexjev-4b-para) Model · Audio text to text ### [DuplexJev-4B-Stress](https://savrn.com/models/duplexjev-4b-stress) [Adventists.ai](https://savrn.com/model-publishers/adventists-ai) Paper: arXiv:2610.02638 · Code: github.com/adventists-ai/duplexjev · Research note: teaching a speech-decision LLM to hear stress DuplexJev-4B-Turn that also hears stress and pauses, for smaller GPUs. "I didn't say she took the money" and "I didn't say she took the money" are the same words; this model tells them apart. It answers which word is stressed, is any word stressed at all, where the speaker pauses, and what the stress implies, in Mandarin and English. Trade-off — read before choosing. At 4B, adding stress costs some spoken-QA accuracy: VoiceBench OBQA / MMSU is 42.0 / 38.2 against 46.4 / 41.7 for 4B-Turn, and qa100 / ZJU-ML are 1 point lower. We tried lower stress weights… Open weights cc-by-nc-4.0 4.2B parameters 40,960 tokens vllm [View model](https://savrn.com/models/duplexjev-4b-stress) Model · Audio text to text ### [DuplexJev-4B-Para-LiveTurn](https://savrn.com/models/duplexjev-4b-para-liveturn) [Adventists.ai](https://savrn.com/model-publishers/adventists-ai) Paper: arXiv:2610.02638 · Code: github.com/adventists-ai/duplexjev · Plugins: pipecat-duplexjev, livekit-plugins-duplexjev The 4B DuplexJev tuned for live end-of-turn: is the user done, right after they go quiet? A turn-taking variant of DuplexJev-4B-Para v1.1, which stays the general 4B. Same recipe as DuplexJev-32B-Para v1.2: the 4B-Para v1.1 mix plus accent data and pause-level turn-taking data, one rank-16 LoRA, 4,000 steps, merged. Qwen3-ASR-0.6B encoder + connector + Qwen3-4B, 4.2 B parameters, bf16, ~10 GB GPU memory. Which 4B to pick. - This one for a voice agent's end-of-turn on a small GPU (LiveKit / Pipecat plugins). - 4B-Para for everything else (stress, non-verbal sounds… Open weights cc-by-nc-4.0 4.2B parameters 40,960 tokens vllm [View model](https://savrn.com/models/duplexjev-4b-para-liveturn) Model · Audio text to text ### [Sori-1B-MCQ](https://savrn.com/models/sori-1b-mcq) [Seonuk Kim](https://savrn.com/model-publishers/snkii) Sori-1B-MCQ — sori (소리) is Korean for sound — listens to a clip and answers typed questions about it: which of these options, how far along this scale, is this true. It returns the answer, a probability for every option you defined, and a confidence — never prose, never a token stream to parse. Code stays in charge: your program defines the options, reads the numbers and decides what to do. It is the decision model of the Sori Agent project at the Human Interface Lab, Seoul National University. Inspired by TypeSafe's System One model, Jev. Jev showed what a model built for software rather than for conversation looks like: a state, typed questions, calibrated probabilities, one call for many… Access requested at publisher other 1.1B parameters transformers [View model](https://savrn.com/models/sori-1b-mcq) Model · Audio text to text ### [MOSS-Transcribe-Diarize-HF](https://savrn.com/models/moss-transcribe-diarize-hf) [Ita Zaporozhets](https://savrn.com/model-publishers/itazap) MOSS-Transcribe-Diarize 0.9B is an end-to-end audio understanding model for long-form multi-speaker transcription, diarization, timestamps, and acoustic event awareness. It supports transcription and diarization across 50+ languages, single-pass inference on audio recordings up to 90 minutes long, and custom hotword prompting for domain-specific terms. Given an audio or video file, the model generates a compact speaker-aware transcript in one pass, including timestamps and anonymous speaker labels such as [S01], [S02], and beyond. 2026-07-22: The subtitle Web UI now supports both Simplified Chinese and English. 2026-07-14: MOSS-Transcribe-Diarize won first place in the 2nd MLC-SLM Challenge… Open weights apache-2.0 909M parameters 131,072 tokens transformers [View model](https://savrn.com/models/moss-transcribe-diarize-hf) ## Qwen [All models and datasets](https://savrn.com/model-publishers/qwen) ## Versions - [0a095220c30b](https://savrn.com/models/qwen2-audio-7b-instruct/versions/0a095220c30b) · current 2026-09-18 ## Explore More - [All audio text to text models](https://savrn.com/models/tasks/audio-text-to-text) - [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-20. - [Qwen website](https://qwen.ai/) - [Hugging Face record](https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct) - [How the hub is built](https://savrn.com/model-hub/methodology)