MiniMax H3-Image from @MiniMax_AI is now available on Merge Gateway. 4K images in under 10 seconds, and edits from up to 16 reference images. One open-weight model handles both. Check it out:

MiniMax just released an image model that generates 16-megapixel images natively, with downloadable weights. H3-Image produces 4096×4096 images without an upscaler, using just six sampling steps. The same model handles generation and editing, with support for up to 16 reference images. One particularly cool example: give it an illustrated fantasy map and it generates a first-person view from the position marked by an arrow, placing the waterfall, bridges and castle in the scene. You can try it through Hailuo or the API, or run it yourself with tools including ComfyUI and Diffusers. The weights use MiniMax's community license, and the full model needs substantial memory.

MiniMax Code now has a Python SDK. pip install minimax-code-sdk Async and sync APIs, multi-turn threads, streaming events, structured errors. Drop MCode into your scripts, evals and data pipelines.

MiniMax H3 Highspeed and H3 Lite from @MiniMax_AI are now available on Merge Gateway. H3 Lite costs half what H3 Max does at 768p ($0.04/sec) and $0.025/sec at 480p, with 5-second clips back in 8 to 9 seconds. H3 Highspeed is 25% under H3 Max at $0.06/sec. Check it out:

MiniMax H3 (Hailuo 3) video on 2× DGX Spark 🎬 • 5 s 1344×768 clip with audio: 643 s (one Spark: 1399 s) • 10 s: 1690 s, 15 s: 3080 s (one Spark: out of memory) • 20 s: 6740 s, beyond the model's 15 s spec and still coherent to the last frame (clip below, one continuous generation, seed 1101, 20 steps) • Text-encoder TP2 + Ulysses SP2 over one ConnectX-7 cable (RoCE, 19.5 GB/s) • Online FP8 DiT, cuDNN attention, 24 fps, native audio • Frames leave the GPU as uint8, VAE tiles decoded on both GPUs: same bytes, 15 s fits with 25 GB to spare on each host • H3-Base open weights (768p), vLLM-Omni, no Ray and no changes to the vLLM-Omni tree vllm serve cannot span two hosts, so the recipe adds a cross-host executor: rank 1 runs on the second Spark and is driven over the link through vLLM's own message queues, behind the unchanged single-node API. Quality checks, memory traces and the open levers are in the repo. Full recipe, measurements and credits: