MOSS-Image: 5B open image model hits 1.3M downloads in four days
neurocartel · reddit · 2026-10-25
- OpenMOSS released MOSS-Image, a 5B open text-to-image model. Per the team and Chinese media (Pandaily, QbitAI): 1.3M downloads in four days, 1,400 GitHub stars, top of HF text-to-image trending.
- Caveat: the download count aggregates across several repos (incl. Comfy-Org on HF), and all figures are team-reported, not independently verified.
- Community momentum: 8-step and 4-step builds already exist (official setting is 50 steps), plus INT8, GGUF, MLX builds and style LoRAs.
- What users want from it: dense Chinese/English text, structured JSON layouts with bounding boxes, prompts up to 4096 tokens.
- Architecture: a 5B diffusion transformer trained from scratch, frozen Qwen3.5-9B text encoder, VAE from FLUX.2. Ostris speculates the more natural look may come from dropping AI-generated images from training data (his guess).
- Testers split: praise for realism and detail, but some styles come out shaky.
More from Models
- Dev: DeepSeek is very hard to beat on easy tasks, 10x cheaper than Sonnet — bindureddy · 2026-10-25
- OpenAI's rumored 6.1 Terra points to demand for smaller, cheaper models — haider1 · 2026-10-25
- SRE Bench: GPT-6 scores 57% at binary reverse engineering, DeepSeek models under 2% — teortaxesTex · 2026-10-25
- DeepMind-backed benchmark: best AI agent solves 49/50 unpublished research-level physics problems — rohanpaul_ai · 2026-10-25
- Why DeepSeek's Flash models dazzle but Pro versions disappoint: a training-signal theory — teortaxesTex · 2026-10-25
- Pro user: 62,500 Codex credits ($2,500) lock to smallest models the day your plan ends — lanqo88 · 2026-10-25