Gemma 4 12B:无编码器多模态,原生音视频理解,16GB 笔记本可跑

Gemma 4 12B: Encoder-Free Multimodal AI with Native Audio & Video, Runs on 16GB Laptop

Tech-News #Gemma#Google DeepMind#多模态#音频理解#视频理解#开源模型#本地部署#LLM
🇨🇳 中文

第一个原生四模态中等规模开源模型

2026 年 6 月 3 日,Google DeepMind 正式发布 Gemma 4 12B Unified——Gemma 系列首个将文本、图像、音频、视频统一纳入同一 LLM backbone 处理的开源模型,参数量 11.95B,256K 上下文窗口,Apache 2.0 协议,16GB 显存的笔记本或 Apple Silicon Mac 即可本地运行。

这不只是参数规模上的更新,而是架构上的根本性改变:彻底去掉了独立的视觉编码器和音频编码器

核心架构:为什么去掉编码器

此前 Gemma 系列中等规模模型的架构是:LLM + 550M 视觉编码器 + 300M 音频编码器。三个独立模块增加了延迟、内存开销,也让跨模态交互变得间接。

Gemma 4 12B Unified 用两个极轻量的投影层替代:

  • 视觉嵌入(35M 参数):原始图像切成 48×48 像素 patch,通过单次矩阵乘法投影成 token,位置编码采用”分解坐标查询”方案
  • 音频嵌入:原始 16 kHz 音频切成 40ms 帧,线性投影到与文本 token 相同的嵌入空间,无需 Conformer 层

两种模态的 token 与文本 token 在同一 transformer 中统一处理,消除了模态转换的信息瓶颈。结果是性能接近参数量两倍以上的 26B MoE 模型,但内存占用不到其一半。

音频理解能力

能力说明
自动语音识别(ASR)无需外部 ASR 管线,直接输入音频
说话人区分(Diarization)识别”谁在说话”
语音转译文本跨语言翻译,CoVoST 基准得分 38.5%
多语言预训练覆盖 140+ 语言,指令微调支持 35+

使用限制:单次音频输入最长 30 秒。模态顺序建议:音频放在文本之后输入。

快速调用示例:

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "请用中文转录以下音频:"},
            {"type": "audio", "audio": "https://example.com/speech.wav"},
        ]
    }
]

视频理解能力

视频处理本质上是”带时间轴的帧序列 + 可选同步音频”。Gemma 4 12B 的处理方式:

  • 按 1 FPS 采样帧,每帧分配 70 token 的视觉 token 预算(适合视频的低预算档)
  • 单次视频输入最长 60 秒(最多 60 帧 × 70 token = 4,200 个视觉 token)
  • 支持音视频同步输入,可同时理解画面内容与对话

官方演示:对一段 5 分钟的 Google I/O 主题演讲片段,模型以 313 帧 + 音频的方式输入,能正确回答”某个功能在哪个时间点被介绍”等细粒度问题。

messages = [
    {
        "role": "user",
        "content": [
            {"type": "video", "video": "https://example.com/keynote.mp4"},
            {"type": "text", "text": "总结这段视频的主要内容,并列出每个功能演示的时间点。"}
        ]
    }
]

主要基准评测

基准得分说明
GPQA Diamond78.8%博士级科学推理
AIME 2026(无工具)77.5%数学竞赛题
MMLU Pro77.2%通用知识
LiveCodeBench v672.0%代码生成
Codeforces ELO1659竞技编程
MATH-Vision79.7%视觉数学
MMMU Pro(视觉)69.1%多模态理解
MRCR v2(128K 长上下文)43.4%超长上下文召回
CoVoST(音频翻译)38.5%语音翻译
FLEURS(ASR,越低越好)0.069 WER语音识别错误率

Google 内部测试:在 Google AI Edge Eloquent 应用中部署后,整体质量提升 60%+

模型系列对比

Gemma 4 共有五个规模档:

模型总参数上下文音频视频
E2B5.1B128K
E4B8B128K
12B Unified11.95B256K
31B30.7B256K
26B A4B MoE25.2B(激活 3.8B)256K

注意:31B 和 26B MoE 不支持音频输入,音视频需求应优先选 12B Unified 或 E2B/E4B。

模型地址

版本HuggingFace 链接
基础模型google/gemma-4-12B
指令微调版google/gemma-4-12B-it
助手版google/gemma-4-12B-it-assistant
Unsloth(GGUF 量化)unsloth/gemma-4-12b-it-GGUF
Unsloth(原格式)unsloth/gemma-4-12b

本地部署

安装依赖:

pip install -U transformers torch accelerate

基础调用模板(推荐参数):

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "google/gemma-4-12B-it"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, dtype="auto", device_map="auto"
)

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,  # 改为 True 开启推理思考模式
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=1.0,
    top_p=0.95,
    top_k=64,
)

图像 token 预算选择(影响精度与速度):

token 预算适用场景
70 / 140视频帧、图像分类、图像描述
280 / 560通用图像理解
1120OCR、文档解析、小字体识别

兼容的推理框架: llama.cpp、MLX(Apple Silicon)、vLLM、Ollama、SGLang、Unsloth、LM Studio

Apple Silicon 一键体验: Google AI Edge Gallery 或 Eloquent 应用(macOS)

为什么值得关注

Gemma 4 12B Unified 的意义在于三个”第一次并发”:

  1. 开源 + 中等规模 + 原生四模态——此前的多模态开源模型要么参数更大,要么不支持原生音频,要么需要多个独立模块拼接
  2. 真正的消费级硬件可跑——16GB 统一内存(M3 Pro MacBook Pro 或 RTX 4080 笔记本均可),不再是”开源但需要 A100”
  3. Apache 2.0 + 商业可用——没有 NC 限制,可以直接用于产品构建

对于想做本地音视频 AI 应用的开发者:Gemma 4 12B 是目前开源生态里硬件门槛最低、模态最完整的选择。


资源链接:

🇬🇧 English

The First Mid-Size Encoder-Free Four-Modality Open Model

On June 3, 2026, Google DeepMind released Gemma 4 12B Unified — the first open model in the Gemma family to process text, images, audio, and video through a single unified LLM backbone. At 11.95B parameters with a 256K context window, Apache 2.0 licensed, it runs locally on a 16GB VRAM laptop or Apple Silicon Mac.

This isn’t just a parameter count update. It’s a fundamental architectural change: complete elimination of separate vision and audio encoders.

Architecture: Why Remove the Encoders

Previous mid-size Gemma models used three separate modules: LLM + 550M vision encoder + 300M audio encoder. This introduced latency, memory overhead, and indirect cross-modal interaction.

Gemma 4 12B Unified replaces them with two ultra-lightweight projection layers:

  • Vision embedder (35M params): Raw images split into 48×48 pixel patches, projected via a single matrix multiplication with “factorized coordinate lookup” for position encoding
  • Audio embedder: Raw 16 kHz audio sliced into 40ms frames, linearly projected into the same embedding space as text tokens — no Conformer layers needed

Both modalities’ tokens are processed alongside text tokens in a unified transformer, eliminating the information bottleneck of modality conversion. The result: performance approaching the 26B MoE model at less than half the memory footprint.

Audio Understanding

CapabilityDescription
ASRDirect audio input, no external ASR pipeline
Speaker DiarizationIdentifies “who is speaking”
Speech-to-TranslationCross-language translation; CoVoST score: 38.5%
MultilingualPre-trained on 140+ languages, instruction-tuned on 35+

Limits: Maximum 30 seconds of audio per input. Recommended order: audio placed after text.

Video Understanding

Video processing treats video as a timestamped frame sequence with optional synchronized audio:

  • Sampled at 1 FPS, with 70 tokens per frame (low-budget tier suited for video)
  • Maximum 60 seconds per video input (60 frames × 70 tokens = 4,200 visual tokens)
  • Supports synchronized audio+video input for joint understanding of visuals and dialogue

Official demo: a 5-minute Google I/O keynote segment was input as 313 frames + audio, enabling fine-grained Q&A like “at what timestamp was feature X introduced.”

Key Benchmarks

BenchmarkScoreDomain
GPQA Diamond78.8%Graduate-level science reasoning
AIME 2026 (no tools)77.5%Math competition
MMLU Pro77.2%General knowledge
LiveCodeBench v672.0%Code generation
Codeforces ELO1659Competitive programming
MATH-Vision79.7%Visual math
MMMU Pro (Vision)69.1%Multimodal understanding
MRCR v2 (128K long context)43.4%Long-context recall
CoVoST (Audio)38.5%Speech translation
FLEURS (ASR, lower is better)0.069 WERSpeech recognition error rate

Internal testing: 60%+ quality improvement in Google AI Edge Eloquent app deployment.

Model Variants

ModelTotal ParamsContextAudioVideo
E2B5.1B128K
E4B8B128K
12B Unified11.95B256K
31B30.7B256K
26B A4B MoE25.2B (3.8B active)256K

Note: 31B and 26B MoE do not support audio input. For audio+video tasks, use 12B Unified or E2B/E4B.

VersionHuggingFace
Base modelgoogle/gemma-4-12B
Instruction-tunedgoogle/gemma-4-12B-it
Assistantgoogle/gemma-4-12B-it-assistant
Unsloth GGUFunsloth/gemma-4-12b-it-GGUF

Why It Matters

Gemma 4 12B Unified achieves three “firsts simultaneously”:

  1. Open source + mid-size + native four-modality — prior multimodal open models were either much larger, lacked native audio, or required multiple stitched modules
  2. Genuinely consumer-hardware capable — 16GB unified memory (M3 Pro MacBook Pro or RTX 4080 laptop), not “open source but needs an A100”
  3. Apache 2.0 + commercial use — no NC restrictions, build products directly

For developers building local audio/video AI applications: Gemma 4 12B is currently the lowest-barrier, most modality-complete option in the open-source ecosystem.


Resources:

💬 评论与讨论

使用 GitHub 账号登录后发表评论