We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8
Comments on “We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Video Rebirth released HyperFlow H3 Flash, a pruned, INT8-quantized, few-step version of MiniMax-H3 that generates video with synchronized stereo audio. It has 13.7B–13.9B parameters versus about 31B, samples in 4–8 steps without classifier-free guidance, and ships GGUF builds (Q4_K_M is 9.1 GB) for 24 GB and 32 GB consumer GPUs. Its own timings are 8.7 s for a 5.2 s clip on 4× B200 and 92 s on one RTX 5090; these and the real-time claim are unverified. Weights exclude US, EU, UK and Korean use without MiniMax's authorization.
Why it ranked
No. 10 of 51 with a score of 0.64. The panel rated it highest on technical depth (0.94) and new information (0.93), lowest on audience breadth (0.50). Placed No. 10, No. 11 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Major open-weight video+audio diffusion model release with detailed technical specs, benchmarks, and quantization methods; limited independent validation and vendor-heavy promotional framing reduce substance somewhat.
- AI judgeJudge 3gpt-5.6-luna
- A highly technical model release with reproducible benchmarks, quantization and distributed-inference details, runnable code, and substantial consumer-GPU guidance. It is broadly relevant to generative-media and inference practitioners, with modest licensing and promotional caveats.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.64total under rubric v2.3
- Significance0.60
- Reader interest0.73
- Audience breadth0.50
- Practical relevance0.78
- Substantiated0.84
- Technical depth0.94
- New information0.93
- Counts against: Promotional0.17
- Counts against: Overclaim0.27
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.63 | 0.45 | 0.72 | 0.60 | 0.27 | +0.24 | +0.144 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.68 | 0.60 | 0.91 | 0.73 | 0.31 | +0.26 | +0.190 |
| Audience breadthHow many people who build technology does it concern? | 0.47 | 0.40 | 0.62 | 0.50 | 0.22 | +0.13 | +0.065 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.75 | 0.70 | 0.88 | 0.78 | 0.18 | +0.22 | +0.171 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.98 | 0.60 | 0.93 | 0.84 | 0.38 | +0.05 | +0.042 |
| Technical depthHow technically substantive is it? | 1.00 | 0.85 | 0.98 | 0.94 | 0.15 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.90 | 0.90 | 0.98 | 0.93 | 0.08 | +0.05 | +0.046 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.08 | 0.35 | 0.08 | 0.17 | 0.27 | −0.19 | −0.032 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.28 | 0.40 | 0.12 | 0.27 | 0.28 | −0.11 | −0.029 |
| Sum of contributions | +0.643 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.643 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-08/10/"><img src="https://rubric.hrih.si/badge/2026-11-08/10.svg" alt="Ranked #10 on Rubric, 8 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 2 matches source · 5 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
HyperFlow H3 Flash has 13.7B (fl2va) or 13.9B (ref2va) parameters, less than half of the roughly 31B-parameter MiniMax-H3 transformer.
The cited source says the same thing.
The article says “Model size: 13.7B (fl2va) or 13.9B (ref2va) parameters, less than half of the ~31B MiniMax-H3 transformer”
The source says “H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer, with approximately 13B parameters residing in AdaLN-related branches”
Source cited by the article: huggingface.co/MiniMaxAI/MiniMax-H3
No source givenBenchmark result
For a 5.2 s clip at 6 steps, HyperFlow H3 Flash takes 8.7 s on 4× B200, 36 s on 1× H200, and 92 s on 1× RTX 5090.
The article gives no link or citation for it.
The article says “Speed (5.2 s clip, 6 steps): 8.7 s on 4× B200, 36 s on 1× H200, 92 s on 1× RTX 5090”
The source says No source to compare with.
Couldn’t checkFirst, only or fastest
NVIDIA SOL-H3 runs HyperFlow H3 Flash in real time on 4× B-series GPUs and supports the model from day 0.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “NVIDIA SOL-H3 runs HyperFlow H3 Flash in real time on 4× B-series GPUs”
The source says No passage could be compared.
Source cited by the article: github.com/NVlabs/Sana/tree/sol-engine/models/minimax_h3/HyperFlow-Flash
Matches sourceNumber
The Q4_K_M GGUF quantization is 9.1 GB (Q8_0 is 14.9 GB), and the model runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM.
The cited source says the same thing.
The article says “Minimum footprint: Q4_K_M is 9.1 GB (Q8_0 is 14.9 GB). It runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM”
The source says “Minimum footprint: Q4_K_M is 9.1 GB (Q8_0 is 14.9 GB). It runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM”
Source cited by the article: huggingface.co/VideoRebirthAI/hyperflow-h3-flash
No source givenCausal conclusion
HyperFlow H3 Flash enables 4-to-8-step sampling without CFG through two-time (t, r) conditioning.
The article gives no link or citation for it.
The article says “Every token receives the current noise level t and the step endpoint r. This enables 4-to-8-step sampling without CFG”
The source says No source to compare with.
No source givenNumber
The INT8 transformer is 10–20% faster than GGUF, while GGUF needs less than half the system RAM.
The article gives no link or citation for it.
The article says “The INT8 transformer is 10–20% faster. GGUF needs less than half the system RAM”
The source says No source to compare with.
No source givenNumber
The 'stream' offload mode is only about 2% slower than 'none' on the same GPU.
The article gives no link or citation for it.
The article says “stream is only ~2% slower than none on the same GPU”
The source says No source to compare with.
No source givenCausal conclusion
FSDP2 sharding of the text encoder cuts 4-GPU peak VRAM from 91 GiB to 59 GiB per GPU.
The article gives no link or citation for it.
The article says “FSDP2 shards the 50 GiB text encoder across GPUs. This cuts 4-GPU peak VRAM from 91 GiB to 59 GiB/GPU”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 60 pts, front page
Recorded for context only, checked 8 Nov 2026, 05:46 UTC. It played no part in the ranking.