LongCap-35B-A3B-FP8

LongCap-35B-A3B-FP8

Turns recordings up to about two hours long into a timeline of speech, music and sound events with start and end times, and also transcribes speech with speaker labels. FP8 weights for a single H100 80GB GPU.

Made by JinchuanTian on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: LongCap listens to a recording, from a few seconds to about two hours, and writes a timeline of what it hears. Each speech, music or sound event gets a start time, an end time and a structured set of attributes. The same checkpoint can also transcribe speech and label who is speaking.

This repository holds the FP8 version, sized for one H100 80GB GPU. The model has about 35.3 billion parameters. A BF16 version also exists.

Who it suits: Audio and media developers, researchers and archivists with a large GPU who need detailed, timestamped descriptions of long recordings. It reads English and Chinese.

Industry | Who uses it | How they use it | Example
Media | Archive and post-production teams | Index long recordings by what happens and when | Marking speech, music and sound cues in a film
Research | Audio researchers | Study long-form audio captioning | Comparing caption quality over two hours of audio
Accessibility | Caption developers | Produce descriptions beyond plain transcripts | Describing sounds along with dialogue
Meetings | Transcription developers | Transcribe speech with speaker labels | Writing a speaker-tagged transcript of a long call

Anyone without an H100-class GPU can skip it for now, since the card sizes the FP8 version for one H100 80GB.

Where to use it:

• A server or workstation with an H100 80GB GPU, using the longcap library.
• The project's code repository and quickstart, which the card points to.
• Windows, macOS or Linux with Python, for downloading the files.

Use case | Who | What they get back
Audio timeline | Media teams | Events with start and end times and structured attributes
Speech transcription | Developers | Transcripts with speaker labels
Long recording captions | Researchers | Captions for recordings up to about two hours in one pass

How to set it up: You need a large GPU and the project's own code.

• An H100 80GB GPU for the FP8 version
• The longcap library from the project's code repository
• Python and the Hugging Face tool, for downloading
• Disk space, since the FP8 weights are about 38.4 GB

1. Install the Hugging Face tool.
2. Download the FP8 files.
3. Follow the quickstart in the project's code repository to run it.

pip install -U huggingface_hub
hf download JinchuanTian/LongCap-35B-A3B-FP8

How to use it: You hand it a recording. It reads the audio piece by piece and logs the events for each piece, while holding on to what came before. The card gives its run steps in the code repository, not in the page text here.

Each event has a type of speech, music or sound_event, a set of attributes and a short caption. The labels inside each event use Chinese words, but spoken words and lyrics stay in the language they were said or sung in.

Tool | What it does
speech event | Carries the transcript and speaker traits
music event | Carries the instruments and tempo
sound_event | Carries the source and distance of the sound

Example: The card ran the model over a 110-minute film in a single pass. Around the 87-minute point the model logged four things: someone's steps fading away, two brief spoken lines and a trumpet solo that signals a change of scene. Each came with times in seconds from the start.

For newcomers and experts: If you are new, think of it as a tool that watches a recording with its ears and writes a detailed log of every voice, tune and noise. Read the quickstart first, since it needs a large GPU.

For experts, the model keeps earlier audio and text in context of up to 1M tokens and uses 62 expert-designed attribute keys across speech, music and sound. Only the mixture-of-experts weights are quantized to FP8, the rest stays BF16, and the card says the BF16 and FP8 versions give closely matching output.

A recording up to about two hours
The model reads it in short chunks
It writes events for each chunk
Earlier audio and text stay in context
You get a timeline with attributes

Title: tcpCapSim on LongCapBench-TAC-Short, higher is better
LongCap: 64.6%
Qwen3.5-Omni-Plus: 52.7%

Good to know: The points below cover limits and costs.

• Output language: attribute names and values are in Chinese, and a key file gives their English names.
• Hardware: the FP8 version targets one H100 80GB, and the BF16 version needs two.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named. A high-end GPU is required.
• Data it reaches: it reads the recordings you give it.
• Not for: setups without a large GPU. A larger model from the paper is not released.

On Hugging Face, 13 November 2026

• Task: audio text to text
• Library: longcap
• Parameters: 35.3B
• Weights: about 35.3 GB in F8_E4M3; about 17.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 65
• Likes: 2
• License: apache-2.0
• Languages: en, zh
• Based on: JinchuanTian/LongCap-35B-A3B
• Tags: longcap, safetensors, qwen3_5, audio, audio-captioning, timestamped-audio-captioning, dense-audio-captioning, long-form-audio, speaker-diarization, automatic-speech-recognition, music, sound-event-detection

Links

• Model: https://huggingface.co/JinchuanTian/LongCap-35B-A3B-FP8
• Files: https://huggingface.co/JinchuanTian/LongCap-35B-A3B-FP8/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download JinchuanTian/LongCap-35B-A3B-FP8

Then follow the usage in its model card. Nothing in this kit runs by itself.