Releases·Model Release
FastVideo releases FastH3 V1, open-weight 4-step sparse distilled MiniMax H3 with 14x speedup on NVIDIA Blackwell
FastVideo released FastH3 V1, an open-weight distilled version of MiniMax H3 that runs in 4 steps using sparse distillation. The team reports a 14x speedup on NVIDIA Blackwell GPUs. The release gives practitioners a faster open option for H3-class models on current-generation hardware.
249 engagement2 sourcesFri, Aug 28, 2026, 09:42 PM
AI Overview
- Distilled from MiniMax H3 via sparse distillation, reducing inference to 4 steps
- Open-weight release, hosted/documented by HaoAI Lab (haoailab.com)
- Reported 14x speedup specifically on NVIDIA Blackwell GPUs
- Single-post thread with low engagement (6); no independent benchmarks yet
Related
Releases·Tool Release
New FP4 attention kernels for B300 achieve up to 1.69x speedup over FlashAttention-4
A new set of FP4 attention kernels optimized for NVIDIA's B300 GPU delivers up to 1.69x speedup compared to FlashAttention-4. The kernels leverage the B300's FP4 tensor core capabilities to accelerate attention computation, a key bottleneck in transformer inference.
A new set of FP4 attention kernels optimized for NVIDIA's B300 GPU delivers up to 1.69x speedup compared to FlashAttention-4. The kernels leverage the B300's FP4 tensor core capabilities to accelerate attention computation, a key bottleneck in transformer inference.
74 engagement1 sourceredditTue, Jul 14, 2026, 12:35 AM
Products·launched
Drehbuchstudio Pro: open-source prompt engine and screenplay studio for cinematic AI video
Drehbuchstudio Pro is an open-source prompt engine and reference anchor generator for multi-scene cinematic AI video generation. It uses LLM-assisted structured prompting to prevent token truncation from multi-line prompts and to anchor character and film-stock consistency across short cut windows. It targets creators using video diffusion models like MiniMax H3, Kling, Runway, and Maestro.
Drehbuchstudio Pro is an open-source prompt engine and reference anchor generator for multi-scene cinematic AI video generation. It uses LLM-assisted structured prompting to prevent token truncation from multi-line prompts and to anchor character and film-stock consistency across short cut windows. It targets creators using video diffusion models like MiniMax H3, Kling, Runway, and Maestro.
1 engagement1 sourceredditSat, Sep 26, 2026, 11:17 AM
Developments·Paper
Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge
A new arXiv paper presents Sol-H3, a full-stack inference pipeline accelerating the 33-billion-parameter MiniMax-H3 open-source video diffusion model on both cloud (NVIDIA GB200) and edge (DGX Spark) hardware. It combines a two-stage cross-resolution generation scheduler with a Recursive Self-Improvement kernel-search loop, achieving up to 30x end-to-end speedup and 20% lower memory.
A new arXiv paper presents Sol-H3, a full-stack inference pipeline accelerating the 33-billion-parameter MiniMax-H3 open-source video diffusion model on both cloud (NVIDIA GB200) and edge (DGX Spark) hardware. It combines a two-stage cross-resolution generation scheduler with a Recursive Self-Improvement kernel-search loop, achieving up to 30x end-to-end speedup and 20% lower memory.
0 engagement1 sourcearxivMon, Sep 28, 2026, 01:15 PM
Developments·Paper
Decomposing Runtime, Kernel, and Quantization Speedups via a Matched FP16 Intermediate: A Hardware-Conditioned Case Study on Four NVIDIA RTX A5000 GPUs
A new paper presents an attribution study that decomposes end-to-end speedups from quantized inference into runtime improvements and kernel/quantization gains. On four NVIDIA RTX A5000 GPUs with NVLink, the full stack achieves 2.58× speedup under greedy decoding, with runtime changes accounting for about two-thirds of the gain on a logarithmic scale.
A new paper presents an attribution study that decomposes end-to-end speedups from quantized inference into runtime improvements and kernel/quantization gains. On four NVIDIA RTX A5000 GPUs with NVLink, the full stack achieves 2.58× speedup under greedy decoding, with runtime changes accounting for about two-thirds of the gain on a logarithmic scale.
0 engagement1 sourcearxivMon, Jul 13, 2026, 10:31 AM
Releases·Tool Release
Gaining 3x inference while running Ollama on iGPU instead of CPU
A blog post demonstrates that running Ollama on an integrated GPU (iGPU) instead of the CPU can yield up to 3x faster inference. The post provides practical guidance for configuring Ollama to leverage the iGPU, which is significant for users without discrete GPUs.
A blog post demonstrates that running Ollama on an integrated GPU (iGPU) instead of the CPU can yield up to 3x faster inference. The post provides practical guidance for configuring Ollama to leverage the iGPU, which is significant for users without discrete GPUs.
1 engagement1 sourceredditSun, Aug 16, 2026, 12:20 PM