MiniMax H3 in four steps

The lead: FastH3 Preview v1

What it is. From FastVideo (Hao AI Lab, with NVIDIA FastGen, Nuva Lab, MiniMax and vLLM), FastH3 Preview v1 is an open-weight, four-step distillation of MiniMax H3 — the open video-and-audio generation model. The base H3 sampler runs 50 denoising steps; FastH3 compresses that into four transformer forwards using DMD2 distillation, while reusing the H3 text encoder, video VAE, audio VAE, tokenizer and scheduler. It generates synchronized video and stereo audio from a text prompt in one pipeline (T2VA), supports variable resolution / aspect ratio / duration in a single checkpoint, and is published as full weights plus a pre-extracted LoRA. The recommended checkpoint is the VSA / Data-Free variant, trained with FastVideo’s Video Sparse Attention (VSA-H3) at 90% sparsity, 64-token tiles, 1300 training steps. License: it inherits the MiniMax H3 Community License (not Apache-2.0), so review the territory and acceptable-use terms before production use.

Evidence. The recommended FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree checkpoint sits at 22,851 downloads / 281 likes since the Aug 27 release — that’s real adoption, not a press release. The full Preview v1 collection ships four full-weight checkpoints plus a LoRA repository (dense and synthetic-data ablations), and FastVideo published the inference code and training recipe alongside. The headline speed figures are vendor-reported and profile-specific: up to 14× speedup on a single NVIDIA Blackwell GPU, and a 15-second 768p clip in under 13 seconds on 8× B200 — those depend on the published hardware, VSA kernel, and warmup protocol. Treat them as claims, not universal guarantees.

Speed / ease. The honest read on the numbers: 14× is real for the Blackwell fast path (the VSA kernel, compilation and VAE parallelism are all part of the measured profile), but the published quick start targets CUDA 13 and 4× B200, and the GPU count must divide H3’s 56 attention heads. On other CUDA systems you add --no-replicated-dit --vsa-kernel triton --no-fa4 and accept different latency. Consumer GPUs, DGX Spark, and Apple MLX recipes are explicitly listed as still-evolving work — so if you’re not on the Blackwell profile, expect slower-than-advertised but still-far-faster-than-50-steps. The weights run on any card with enough memory.

Scope caveat. Preview v1 distills only the text-to-video-and-audio path. First/last-frame conditioning (FL2VA) and reference-to-video (Ref2VA) were not distilled in the released checkpoints — FastVideo lists them as follow-up work. And the team states plainly that preview quality remains below Base H3 on difficult motion, fine detail, and some audio; four steps target minimum latency, and an eight-step quality setting is being evaluated. Dense attention is not a drop-in substitute — the recommended checkpoint needs the VSA path.

Deployment verdict. Incumbent: the open MiniMax H3 base you already run at 50 steps. The delta is 4 steps vs 50 — a genuine 10×+ fewer denoising evaluations, with synchronized video-and-audio preserved, at a preview-quality cost and a T2VA-only scope. For anyone whose H3 pipeline is latency-bound, this is the most meaningful speedup to land on the open stack in months; for anyone who needs first-frame/ref conditioning or maximum quality, it’s not ready. Verdict: Marginal — adopt only if you want sub-realtime open H3 on the Blackwell fast path and can live with preview quality plus T2VA-only scope; wait for the RTX / DGX Spark / MLX recipes before treating it as a universal drop-in.

Models: TimesFM-3 and Spark-X2.5-1.7B

TimesFM-3 — from Google Research, a 330M-parameter zero-shot time-series foundation model that is natively pre-trained for multivariate forecasting, and forecasts the whole horizon in a single forward pass (Contiguous Patch Masking) instead of one patch at a time. It jointly predicts multiple coevolving series and accepts past covariates and past-future (dynamic) covariates — e.g. a planned promotion schedule — with point and quantile (9 quantiles) outputs. Evidence: 123,025 downloads last month / 465 likes — big real adoption for a forecasting model. Benchmarks are vendor-reported: top-ranked on Gift-Eval, FEV-Bench and Time among pre-trained foundation models, beating Chronos-2 and the Toto 2.0 family. It’s genuinely light (0.3B, F32) and runs on a laptop or CPU, with a single-pass decode that cuts latency and error accumulation. One important caveat: the license is the TimesFM Non-Commercial License v1.0 — fine for research and personal use, but it blocks commercial self-hosting. Verdict: Real deployment option for personal/research forecasting; marginal for production — the license blocks commercial local serving, so check it before you ship.

Spark-X2.5-1.7B — the smaller sibling of the Spark-X2.5-4B we covered yesterday, same XHToken / Apache-2.0 family, hybrid-attention architecture and native 1M-token context, 200+ languages. Evidence: 2,301 downloads / 94 likes — clears the gate but modest. It’s a follow-up development on a covered base, and it’s genuinely tiny (a ~1.7B model runs anywhere). Verdict: Marginal — adopt only if you want the smallest possible member of the Spark-X2.5 family; the 4B is the sweet spot and this is the pocket-size option.

Quantizations: one-file H3 turbo, plus a few builds

MATLOWAI / minimax-h3-fused-turbo-int8-convrot is the quant of the day, and it’s a strong one: 48,756 downloads / 92 likes. It’s a single 21 GB ComfyUI diffusion-model file that does MiniMax H3 text/image-to-video and reference-to-video in four steps, with a distilled turbo LoRA (8-step v1.0 @ 1.0) and a motion-smoothing Mystic v2.0 LoRA (@ 0.7) already folded into the weights, and INT8 ConvRot quantization (group 256, per-channel) applied once after the merge on all four heavy Linear weights per block — everything else stays BF16/F32. The base is a pruned FL2VA transformer with a rank-1024 SVD of the (ref2va − fl2va) delta fused in, so one partition serves both first/last-frame and reference conditioning. Measured (not estimated): on a 96 GB RTX PRO 6000, a 10-second 1152×640 clip at 4 steps takes 76 seconds; 8 steps is 103 s; 25 steps 292 s. The single-file bake saves about 21 GB of VRAM vs stacking the LoRAs live (the unpatch backup on 200 patched INT8 layers is the whole model again): 47.8 GB peak baked vs 68.9 GB peak live, same wall time, same take. It needs the SLA sparse-attention node pack and a ROCm/CUDA fork, and res_multistep beats euler for audio at 4 steps. Verdict: Real deployment option for H3 ComfyUI users — one file, 4-step, measured-identical output, ~21 GB VRAM saved; the trade is you give up turning the LoRAs down.

One-liners. orcarouter/GLM-5.3-Flash-Uncensored-FP8 (8,338 dl / 186 likes) — an FP8 build of the covered GLM-5.3-Flash with the safety layer stripped; marginal, adopt only if you want an uncensored FP8 GLM-5.3-Flash. orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF (106,845 dl / 240 likes) — uncensored GGUF of the Qwen4 preview; high adoption, but it’s the same base we already covered with an uncensored pass. Jackrong/Qwopus3.8-27B-Flash-GGUF (10,680 dl / 121 likes) — another Qwen3.8-27B GGUF build with MTP + speculative decoding; marginal, a build not a breakthrough. nvidia/Qwen3.8-Flash-Next-NVFP4 (1,129 dl / 104 likes) — the NVFP4 build, already noted in the Qwen4 architecture post.

Fine-tunes

One fine-tune clears the gate and it’s a follow-up: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF (174,405 dl / 231 likes) — a further DavidAU uncensored Qwen3.8-27B TURBO build (cuts thinking tokens, abliterated, with MTP GGUF), on the same base we covered yesterday. It’s real adoption but it’s another uncensored Qwen3.8-27B variant, not a new capability. Verdict: Marginal — adopt only if you want the fastest uncensored Qwen3.8-27B for creative writing; nothing here beats the model you already run. Every other fine-tune candidate today sits at 0 downloads — brand new, no adoption signal, no headline.

Technical term of the day: step distillation (DMD2)

What it is. Diffusion models don’t generate an image or a video in one shot — they denoise it over many steps, each a full transformer forward. MiniMax H3 uses 50. Step distillation trains a student model to predict the outcome of many denoising steps in a single forward, so the student needs far fewer steps to reach the same result. DMD2 (Distribution Matching Distillation, the method FastH3 uses) trains the student by matching the distribution of its one-step output against the teacher’s multi-step output — it’s the reason FastH3 runs a trained five-point sigma grid that produces exactly four DiT calls instead of fifty. The result: latency drops roughly with the step count, while quality is preserved for the distilled operating point.

Why it matters for local AI. For video-and-audio generation, the bottleneck isn’t just model size — it’s how many times you run the transformer. A 33B H3 at 50 steps is a lot of forwards; at 4 steps it’s a fraction of the compute for the same clip. That’s the difference between waiting 90 seconds and waiting 10, or between fitting on a mid-range card and needing a big one. The same idea is why FastWan, LTX, and the Qwen-Image Lightning LoRAs (8-step) all exist: fewer steps is the cheapest speedup available, because it multiplies across every generation.

The trade-off. Distillation trades fidelity for speed. Fewer steps means difficult motion, fine detail and some audio can trail the base model — FastH3’s own team says exactly that, and is testing an eight-step quality setting. There’s also a scope cost: a distilled student is trained for a specific conditioning path (FastH3 preview v1 is T2VA only), so other modes need their own distilled checkpoint. And the distilled operating point is tied to its scheduler and attention path — swapping to dense attention or a different sampler can break the intended setup. The practical rule: fewer steps is the fastest win and the most fragile one — adopt it where latency is the pain, and keep the base model for the hard cases.

Harness improvements: local agents get one-click

The harness news today is the IFA 2026 local-AI push from NVIDIA, Microsoft and partners — and it lands squarely on the agent stack. Three widely-used agent apps get simplified local model setup on Windows: Hermes Agent (Nous Research) gets one-click local model configuration across RTX and DGX systems — auto-detect the NVIDIA GPU, pick a model and config, and run it through integrated llama.cpp with NVIDIA’s optimizations already in place, no manual downloads or tuning; OpenClaw (the open-agent project, 380K+ GitHub stars) gets a Windows App that simplifies setting up an optimized local model on any RTX GPU with ≥24 GB VRAM; Perplexity Portable Computer brings its local-first agent to RTX GPUs with ≥24 GB VRAM on Linux, with Windows support coming. All built on llama.cpp and incorporating NVIDIA’s latest inference optimizations.

Why it matters. The friction of local agents — pick a model, find a compatible server, dial in quantization, keep it updated — is what has kept “run it on my machine” aspirational for most people. One-click setup removes it, and it does so for the harnesses you already run. The inference gains are concrete: llama.cpp up to 1.9× throughput on a GeForce RTX 5090 (kernel optimizations, better speculative decoding, faster prefill); vLLM 1.2× on RTX PRO 6000 Blackwell and up to 1.4× on two DGX Spark clusters (new XQA attention kernels in FlashInfer). Those flow through LM Studio and Ollama too. And NVIDIA PAIR — a free, open-source Personal AI Router — discovers compatible PCs on your local network and routes independent inference requests to whichever has capacity (works with Ollama and LM Studio, Windows/macOS/Linux, Apple M4+ and RTX 20-series+). The example is telling: ask an agent to sort a cluttered inbox, and it splits the work across subagents while PAIR distributes those jobs across your machines instead of all queuing on one GPU. New compact NVIDIA RTX Spark Windows PCs (Lenovo, Acer) land in October. For a public post: this is the strongest single push toward “local agent on your own hardware, no cloud” we’ve seen, and it’s aimed at the exact harnesses people use.

The takeaway

Today’s story is speed on the stack you already have. FastH3 takes the open H3 video-and-audio model from 50 steps to 4 — a genuine latency win with real adoption (22.8k downloads), preview-quality caveats and a T2VA-only scope; the single-file INT8 turbo build gives ComfyUI users the same 4-step speed with ~21 GB less VRAM. TimesFM-3 shows a small, single-pass multivariate forecaster with big adoption (123k) but a non-commercial license. And the IFA 2026 harness push is the most practical one yet: one-click local models in Hermes, OpenClaw and Perplexity, 1.9× llama.cpp, and PAIR spreading work across the PCs you own. Step distillation is the quiet engine behind all the “N steps instead of 50” releases — the cheapest speedup there is, and the most fragile.

Credits

• FastVideo (Hao AI Lab) — FastH3 Preview v1 (with NVIDIA FastGen, Nuva Lab, MiniMax, vLLM); speed, quality and scope from the release card + fasth3.org guide.
• MATLOWAI — minimax-h3-fused-turbo-int8-convrot (measured timings from the model card).
• Google Research — TimesFM-3 (330M, single-pass, TimesFM Non-Commercial License v1.0).
• XHToken (SparkLLM) — Spark-X2.5-1.7B.
• NVIDIA / Microsoft / partners — IFA 2026 local-AI announcements (Hermes Agent, OpenClaw, Perplexity, llama.cpp, vLLM, NVIDIA PAIR, RTX Spark).

Sources: HuggingFace model cards and the HF API (downloads / likes / community builds verified directly), plus the FastVideo release card, fasth3.org guide, Google Research blog, and NVIDIA IFA 2026 blog. Speeds are measured where a benchmark is cited; where not, they’re labeled estimates. No fabricated numbers.