Greetings devs. Now that Mistral Small 4.0 has officially released, I am looking forward to future Mistral models. Like a Mistral Omni model with TTS, STT, image recognition, and OCR. This could go head to head with Qwen 3 Omni. Maybe use new reasoning technologies. With the new NVIDIA partnership and Qwen shifting gears, experimentation would be cool to see. @finite lake , digging the new Mistral direction with NVIDIA. Latent space reasoning, omnimodal models will be the future. Would be a nice way to end 2026.
#Mistral Omni 26B MoE
4 messages · Page 1 of 1 (latest)
Support ?
Do you need help with something? You can contact support to get help if needed.
Will this use Turboquant tech as well?
Update 2026-04-02: Gemma 4 26BA4 released. It is not fully omnimodal, but goes to show that Google knows the size that works well for 32 GB memory. If Mistral released a model this size with omnimodal tech like Qwen 3 Omni, France will be on top. Multi-token prediction and DeepSeek Engrams would be nice features as well. Also, latent space reasoning would be like the holy grail.
Edit 2025-04-05: Tested Gemma 4's translation capabilities, and Gemma 4 swept Mistral Small 3.2 away. Also, Kimi has a new paper where attention can be used between layers:
https://arxiv.org/abs/2603.15031