Ollama v0.33.3：MLX引擎Gemma4模型新增图像与音频输入支持
v0.33.3: gemma4: image and audio input support
Gemma4多模态能力正式落地，涉及新输入协议与切片逻辑，建议开发者立即升级以启用图像/音频处理功能。
Safetensors gemma4 imports served by the MLX engine now answer image
由 MLX 引擎服务的 Safetensors gemma4 导入现在支持图像
and audio chats. Images run through both vision architectures: the
和音频聊天。图像通过两种视觉架构运行：
transformer tower (26B, 31B, e-series) and the 12B's encoder-free
Transformer 塔（26B、31B、e系列）以及 12B 的无编码器
unified embedder. Audio arrives through the same intake the ollama
统一嵌入器。音频通过 ollama
API already accepts for gemma4 GGUFs — WAV bytes in the images field,
API 已经接受的相同入口进入，用于 gemma4 GGUFs——WAV 字节在 images 字段中，
OpenAI input_audio parts, and /v1/audio/transcriptions uploads — with
OpenAI input_audio 部分，以及 /v1/audio/transcriptions 上传——其中
the e2b/e4b checkpoints running clips through their conformer audio
e2b/e4b 检查点通过其 conformer 音频
encoder and the 12b unified checkpoint embedding the raw waveform
编码器运行剪辑，而 12b 统一检查点直接嵌入原始波形
directly. Clips longer than 30 seconds are split evenly into chunks
时长超过 30 秒的剪辑被均匀分割成最多 30 秒的块，
of at most 30 seconds, cut at pauses, and encoded independently.
在停顿处切割，并独立编码。
Each modality serves only checkpoints that carry it: 26B/31B have no
每种模态仅服务于携带它的检查点：26B/31B 没有
audio config and reject audio input, and checkpoints with an
音频配置并拒绝音频输入，而具有
unrecognized vision architecture still load as text-only models and
无法识别的视觉架构的检查点仍作为纯文本模型加载，并
reject image requests.
拒绝图像请求。
The server previously hid the vision and audio capabilities for
服务器之前隐藏了 vision 和 audio 功能，因为
gemma4 safetensors because the engine served neither. Both
gemma4 safetensors 由于引擎未服务两者。现在
suppressions are removed, and existing imports start advertising the
移除了这两种抑制，现有导入开始宣传这些
capabilities without re-importing since import already records them.
功能，无需重新导入，因为导入已记录它们。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力