Kimi-Audio 2: There's No v2 — What Kimi-Audio-7B Actually Is

Is there a Kimi-Audio 2?

No — there is no "Kimi-Audio 2." As of June 2026, Moonshot AI has shipped only one version of its audio model: Kimi-Audio, with weights published as Kimi-Audio-7B (base) and Kimi-Audio-7B-Instruct (instruction-tuned). It launched in April 2025 and has not had a v2.

The confusion is understandable: Moonshot's text models iterate fast (the K2 line), but the audio model has stayed on its first release and was never folded into K2. So if you searched for "Kimi-Audio 2," the model you actually want is Kimi-Audio-7B-Instruct — everything below covers it.

What Kimi-Audio actually is

Kimi-Audio is an open-source audio foundation model: a single 7B model (initialized from Qwen2.5-7B) that does audio understanding and generation in one framework, including:

• Speech recognition (ASR) — speech-to-text
• Audio question answering and audio captioning
• Speech-emotion and sound-event classification
• End-to-end speech conversation (speech in, speech out)

It takes continuous acoustic features plus discrete semantic tokens as input, and emits text tokens plus discrete audio tokens, with a chunk-wise streaming detokenizer (flow matching) for low-latency generation. It is not a pure TTS or pure ASR tool — it's a multi-task understanding-plus-conversation model. It was trained on over 13 million hours of audio and officially supports English and Chinese. The license is dual — Apache 2.0 for the Qwen2.5-derived code and MIT for the rest — so commercial use is allowed, though you should also check the underlying Qwen2.5 terms.

Can your GPU run it?

Moonshot has never published a VRAM requirement — and that gap is the single most-asked question about the model. Users have asked directly, with no official answer (issue #16).

What the field reports actually show:

• A 24GB RTX 4090 hits CUDA out-of-memory running the official example (issue #39).
• One community guide (a third-party blog, not official) reports a 16GB Mac can't hold the full BF16 model and suggests 32GB+ of memory.

So plan for more than 24GB of GPU memory, or expect to quantize and optimize. Treat the "~23GB / 24GB recommended" figure circulating online as a community estimate, not a spec — it is not in Moonshot's official model card.

Why audio longer than 30 seconds stalls

A recurring, still-unanswered failure: generation freezes at Generating tokens: 1% — no error, no output — once the input audio runs past roughly 30 seconds (issue #157). The model's maximum input length is undocumented (issue #158). Until Moonshot responds, the workaround users fall back on is splitting audio into sub-30-second chunks.

Install Kimi-Audio

Three official paths (from the GitHub README and Hugging Face card):

# 1) From source
git clone https://github.com/MoonshotAI/Kimi-Audio.git
cd Kimi-Audio
git submodule update --init --recursive
pip install -r requirements.txt

# 2) Direct pip install
pip install torch
pip install git+https://github.com/MoonshotAI/Kimi-Audio.git

# 3) Docker
docker pull moonshotai/kimi-audio:v0.1

Heads-up: the requirements.txt / submodule step is the most common place to get stuck — users report dependency and CUDA / torch / transformers version conflicts with no confirmed "known-good" combination (issue #113).

Known rough edges

Beyond VRAM and long audio, the open issues worth knowing before you commit:

• The official audio2audio example errors out with AttributeError: 'NoneType' object has no attribute 'shape' (issue #122).
• Multi-turn conversation crashes as history grows — a tensor-reshape error around turn 8 (issue #120).
• Language instructions aren't always followed — a Chinese-only prompt can still return English (issue #134).
• No speaker diarization — it can't tell speakers apart; the feature request is open (issue #149).

Most of these are open and unanswered by Moonshot, so budget time for trial-and-error.

How Kimi-Audio compares (ASR)

On Moonshot's own benchmarks (technical report, arXiv:2504.18425), Kimi-Audio's word error rate (WER, lower is better) beats Qwen2-Audio and Qwen2.5-Omni across English and Chinese ASR:

These are self-reported numbers from Moonshot's paper, not independent tests. For reference, OpenAI's Whisper large-v3 is often cited around 1.8–2.0 WER on LibriSpeech test-clean per OpenAI's own reporting — a separate source, not part of Moonshot's table.

Sources: official GitHub repo, Hugging Face model card, and the Kimi-Audio Technical Report. Issue numbers link to the project's GitHub tracker. Benchmark figures are Moonshot's self-reported results; VRAM and latency are community reports where noted — Moonshot has published no official hardware spec.