Model Releases

  • black-forest-labs/flux-3-image-klein-comfy — ComfyUI node wrapper for Flux 3 image generation/editing. If you run workflows in ComfyUI, this gives you single-file access to the latest Flux variant without the usual config hassle. 🤖
  • meta-models/Muse-Spark-Open-816B-A42B — Meta’s open-weight multimodal model, ~816B total params across a 42B active subset. Conversational + image understanding. Huge footprint, probably needs serious GPU allocation to run at meaningful throughput.
  • deepseek-ai/DeepSeek-V4.1-Flash-1017 — DeepSeek’s V4.1 Flash, supports both text and image generation. The 8-bit FP8 quantization option is worth noting if you’re optimizing for memory-constrained serving. 📄
  • black-forest-labs/flux-3-image-klein — The diffusers-native release of Flux 3 image generation. Same architecture as the ComfyUI variant above but packaged for standard HF diffusion pipelines. Pick one based on your stack. 📄
  • Qwen/Qwen3.8-27B — Qwen’s 27B conversational model under Apache 2.0. If you need an open-weight alternative with a permissive license and decent performance on Chinese-heavy workloads, this is a solid bet. 🤖
  • facebook/sam3 — Meta’s SAM3 video segmentation model. Adds video-level instance segmentation on top of the original SAM architecture. Useful if you’re doing temporal segmentation tasks, otherwise it’s a neat reference implementation. 🔥

Today’s Synthesis

Build a multimodal serving pipeline using flux-3-image-klein for image generation through the standard HF diffusers interface, paired with DeepSeek-V4.1-Flash-1017 running in 8-bit FP8 quantization for any accompanying text generation needs. The HF diffusers path gives you native integration with existing MLOps tooling, while the FP8 quantization keeps the text side from consuming all your VRAM on shared GPUs. This gives you a single-resource multimodal stack that handles both image generation and captioning without dedicated hardware per task. If your team already runs diffusers pipelines, this is the lowest-friction upgrade path to get production-grade Flux 3 up and serving alongside a capable text model — no ComfyUI config required, just swap in the quantized variant and route outputs through your existing serving layer.