GigaChat Audio 10B (A1.8B)

GigaChat Audio 10B is an audio-native LLM built on top of the GigaChat 3.1 Lightning text model. A Conformer speech encoder and a modality adapter feed audio embeddings directly into a Mixture-of-Experts decoder, so the model keeps the text quality of its base while adding speech understanding.

Capabilities: audio question answering and classification, temporal grounding (localization in long audio, timestamped event descriptions, audio summarization with timestamps), tool-use, and text-only tasks.

The temporal grounding skills are trained on TimeGround-1M — a purpose-built dataset of long-form audio paired with time-aligned annotations.

Evaluation

1. Core audio tasks vs open models

2. Timing tasks (detailed)

Full timing metrics across length buckets

TL — temporal localization: find when something is discussed (mIoU vs the reference span). TD — timestamped descriptions of audio events (overall grade 0–5). SUM — long-audio summarization with timestamps; single score = mean of factual accuracy, timing structure and audio coverage with timestamps.

Metric	Bucket	GigaChat Audio (10B-A1.8B)	Voxtral (3B)	Phi-4 (4B)	Qwen3-Omni (30B-A3B)
TL mIoU ↑	≤10m	40.3	3.4	0.2	12.9
TL mIoU ↑	10–20m	46.8	0.0	0.2	0.0
TL mIoU ↑	20–60m	48.3	0.1	0.2	0.1
TL mIoU ↑	60–120m	48.9	0.9	3.7	4.6
TL mIoU ↑	AMI meeting	30.3	0.0	0.0	0.0
TD overall ↑ (0–5)	≤10m	3.45	2.89	2.62	3.23
TD overall ↑ (0–5)	10–20m	3.21	2.33	2.07	2.49
TD overall ↑ (0–5)	20–60m	3.27	2.15	1.95	2.17
TD overall ↑ (0–5)	60–120m	3.26	1.45	1.21	1.84
SUM overall ↑	≤10m	71.6	64.1	26.2	65.7
SUM overall ↑	10–20m	71.4	58.9	23.0	54.8
SUM overall ↑	20–60m	67.9	50.2	21.4	43.8
SUM overall ↑	60–120m	55.4	40.1	9.8	17.3

3. Text quality vs the text base model

Adding the audio modality shifts text quality: some benchmarks regress (MMLU-Pro, IFEval-ru, BBH), others improve (RuBQ, GPQA Diamond).

---

Quickstart (Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

model_name = "ai-sage/GigaChat3.1-Audio-10B-A1.8B"
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda:0",
)

messages = [{"role": "user", "content": [
    {"type": "audio", "path": "90min_lecture.wav"},
    {"type": "text", "text": "When does the speaker mention Star Wars midi-chlorians?"},
]}]
inputs = processor.prepare_for_inference(messages, device=model.device)
output_ids = model.generate(**inputs, max_new_tokens=256)
answer_ids = output_ids[0, inputs["input_ids"].shape[1]:]
print(processor.decode(answer_ids))
# > ... in the interval 01:27:46 to 01:27:53. In this segment they explain ...

vLLM (single GPU, native multimodal)

Both encoder and decoder run inside vLLM. compilation_config disables torch.compile (needed for the Conformer tower); max_num_batched_tokens sizes the audio encoder cache (~90 min is 33k tokens):

import librosa
import os
from transformers import AutoProcessor
from vllm import LLM, SamplingParams

os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")

model_name = "ai-sage/GigaChat3.1-Audio-10B-A1.8B"


def main():
    processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
    llm = LLM(
        model=model_name,
        trust_remote_code=True,
        max_model_len=65536,
        max_num_batched_tokens=65536,
        limit_mm_per_prompt={"audio": 1},
        compilation_config={"mode": 0, "cudagraph_mode": "FULL"},
    )

    messages = [{"role": "user", "content": [
        {"type": "audio", "path": "90min_lecture.wav"},
        {"type": "text", "text": "When does the speaker mention Star Wars midi-chlorians?"},
    ]}]
    text, audio_paths = processor.render_prompt(messages)
    wav, _ = librosa.load(audio_paths[0], sr=16000, mono=True)
    out = llm.generate(
        {"prompt": text, "multi_modal_data": {"audio": [(wav, 16000)]}},
        SamplingParams(temperature=0.0, max_tokens=256),
    )
    print(processor.decode(out[0].outputs[0].token_ids))
    # > ... in the interval 01:27:46 to 01:27:53. In this segment they explain ...


if __name__ == "__main__":
    main()

Throughput

Single-request decode speed on one H100 (vLLM 0.18.0, bf16, greedy), by amount of audio held in context:

Audio features (encoder + adapter)

model.encode_audio returns the LLM-space audio embeddings:

spec = processor.feature_extractor.process("audio.wav")   # (frames, 64) log-mel
specs = spec.unsqueeze(0).to(model.device, torch.bfloat16)
lengths = torch.tensor([spec.shape[0]], device=model.device)
feats, feat_lengths = model.encode_audio(specs, lengths)   # (1, tokens, hidden)

Recommended versions

• torch 2.10.0, torchaudio 2.10.0
• transformers 4.57.6
• vllm 0.18.0
• flash-attn 2.8.3

Citation

If you use GigaChat Audio in your research, please cite:

@misc{kutsakov2026_gigachataudio,
  title         = {{GigaChat Audio}: Time-aware Large Audio Language Model},
  author        = {Kutsakov, Aleksandr and
                   Sadovina, Mariia and
                   Gospodinov, Georgii and
                   Maximenko, Alexandr and
                   Kutuzov, Oleg and
                   Bogomolov, Pavel and
                   Minkin, Fyodor},
  year          = {2026},
  eprint        = {2607.10387},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2607.10387},
  note          = {Accepted to Interspeech 2026}
}