
FastFrames Video Generation vs MiniMax H3: Three Boards Against a Stopwatch
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 87 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 118 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 187 tok/s
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 143 tok/s
There is a version of this comparison that puts two video models side by side and picks the faster one, and it would take about four hundred words. It would also be wrong, because MiniMax H3 is a model you can call by the second and FastFrames Video Generation is a set of weights you download and run on hardware you own. H3 launched on 31 July 2026 as an omni-modal system that reads text, image, video and audio references as one context, generates native stereo sound, and holds second place on the Artificial Analysis image-to-video board at 1,181 Elo over 7,284 votes as read on 8 October 2026. FastFrames arrived on 16 November as two checkpoints, 1.3B and 14B, distilled from Wan2.1 to produce a 480p frame in 36 milliseconds — and it appears on no independent board at all. What separates them is not quality. It is that one of these is a product with a public scoreboard record and the other is a paper with a stopwatch.
Three boards, one model, and what each board is actually asking
H3's positions are worth stating precisely, because the number people quote is usually the wrong one.
• Image-to-video — second at 1,181 Elo over 7,284 votes, behind MiniMax H3 Max at 1,195 and ahead of Vidu Q4 Preview at 1,179. Sixteen points cover the top three, which is inside the intervals; that is a tie dressed as a podium.
• Text-to-video — fourth at 1,137 Elo over 8,315 votes. The model's weaker board, and the one where a 480p silent research model would be competing if it had an entry.
• Video editing — fourth at 1,119 Elo over 7,023 votes, on a board that ranks models on editing a clip from a text instruction with the audio preserved.
Three appearances mean three different capabilities were submitted and measured: generating from a still, generating from text, and transforming footage that already exists. The third is the one that matters most for this comparison, because it requires accepting video as an input, and FastFrames accepts nothing but text.
The pricing sits on our own model card at $0.08 per second at 768p and $0.13 per second at 2K, which is $4.80 and $7.80 per minute — the same pair of figures that appear on the arena's price column, and the same model at two resolution tiers rather than a premium and a cheap version. A 5-second clip at 768p costs $0.40.
What 36 milliseconds does not include
FastFrames' speed is real and it is not a general-purpose speed, and the architecture explains why in a way that a spec list does not.
Both checkpoints are autoregressive video diffusion models distilled from Wan2.1 through a recipe called Causal Forcing++. The authors left the diffusion transformer mostly alone and replaced the two things around it that were never the point. The VAE decoder — about 147 milliseconds a frame in the Causal Forcing++ baseline — becomes a 63.87 million-parameter pixel decoder that emits frames directly and lands near 1 millisecond. The KV-cache update becomes a small refinement network costing 2 milliseconds on the 1.3B model and 6 on the 14B. The transformer is frozen during the final stage.
On the small model that takes 271.8 milliseconds a frame down to 35.8. On one B200, FastFrames-S runs about 28 requests a second and produces a 5.06-second clip in 2.9 seconds. The 14B holds 45 milliseconds only across eight B200s; on a single GPU it degrades to 158 milliseconds a frame, which is slower than watching the result. At 720p it is 127 milliseconds on the small model, and the paper puts the H100 penalty at 25 to 50 percent.
What the number does not include is everything H3 spends its time on. FastFrames produces no audio. It conditions on text only — no start frame, no end frame, no reference image, no clip to extend or restyle. It generates 81 frames at 832×480 and 16 frames per second, and it chains blocks rather than offering keyframe or camera control. Three of the four input modalities H3 accepts do not exist in FastFrames' inference path.

Why the speed is still worth something even so
It would be easy to write FastFrames off after the previous section, and that would be the wrong read. The number is not a marketing figure; it changes what kind of product is possible.
At 36 milliseconds a frame on one B200, a generation finishes before you can watch it. That is the threshold where a preview loop stops being a batch job and becomes interactive: you type, you see, you adjust, and the machine keeps up with the adjustment. A 14B model at 45 milliseconds across a node gets there in the batched-throughput sense, but the single-stream latency that a person actually feels is the small model's property. Nothing in H3's published surface addresses that use case, because H3 is metered in seconds of finished footage rather than in frames produced per second of wall clock — the two are not the same measurement and the rate card does not let you derive one from the other.
There is a second-order consequence that the paper makes explicit. The pixel decoder is trained per base model, but the principle — freeze the transformer, learn the plumbing — has already been ported to third-party systems in the paper's own experiments, taking LingBot-World from 1,085 to 52 milliseconds a frame on the 1.3B variant and from 3,862 to 233 on the 14B. That is a result about a technique, not a product, and it is the part of this release with the widest reach.
The honest state of the evidence on each side
This is where the matchup stops being arguable in either direction, and it is worth being blunt about both halves.
Everything FastFrames claims about quality comes from its own authors. The VBench totals in Table 1 are 84.23 for FastFrames-S and 84.44 for FastFrames-L, against 84.11 for the baseline at 272 milliseconds and 84.65 for Wan2.2 14.3B at 117,734 milliseconds. Nobody independent has reproduced them. VBench rewards the temporal smoothness autoregressive models are built to produce, and the entire spread on Apple's own table is about two and a half points while the latency spread across the same models is four orders of magnitude. Read the table as a claim about parity, not a measurement of it.
H3's board positions carry sample counts and confidence intervals, which is a stronger form of evidence and also a narrower one. A preference board tells you which output people chose when shown two clips; it does not tell you what either clip cost to produce, whether the sound was in sync, or whether the model would survive a 40-request burst at 2K. The sixteen-point spread at the top of the image-to-video board is a good reminder that these numbers rank perceptions, not capabilities, and that any page declaring a winner from a sixteen-point gap on ±10 intervals is reading noise.
So the comparison comes down to this: one model has been voted on by thousands of people and priced per second, and the other has a stopwatch, a licence file and a paper. Those are different kinds of evidence and neither substitutes for the other.
Two procurement paths that do not resemble each other
Buying time on H3 and running FastFrames involve almost no shared steps, which is the practical reason they rarely end up in the same decision.
• H3 — a per-second rate on a hosted endpoint, at $0.08 at 768p and $0.13 at 2K, on the vendor's own API and through several third-party platforms. No hardware, no licence review, no CUDA. We route it as minimax/minimax-h3 on the public model API, billed at the vendor's list rate passed through with nothing added, callable on the same OpenAI-compatible endpoint as the language models in the catalogue.
• FastFrames — Linux x86-64, an NVIDIA GPU, CUDA 12.8, then uv sync and a scripts/generate.py invocation with a checkpoint directory and a seed. The weights carry Apple's AMLR terms; the LingBot-World v2 demo components in the same repository are CC BY-NC-SA 4.0, which is non-commercial; the rest sits under Apple's Sample Code Licence. Apple also published every training-stage checkpoint for the large model, which is unusual and makes the release reproducible in a way most open-weight drops are not.
• The hidden cost on both sides — H3's per-second meter runs whether or not the take was good, and a rejected take bills the same as an accepted one. FastFrames' GPU meter runs the same way, but at a marginal rate that is a rounding error if the hardware is already amortised and a large fixed cost if it is not.
One practical consequence of pass-through pricing on our side: when MiniMax moves a per-second rate, that movement is visible on the route the same day rather than at contract renewal. FastFrames has no rate to move.
What each one is for, stated as a sentence you can act on
Take MiniMax H3 when the input is anything other than a text prompt, when the deliverable needs sound, or when a fifteen-second clip with multi-turn extension is closer to your shot than 81 frames of 480p. It is the only one of the two that can edit footage, the only one that generates audio, and the only one whose quality claim has been tested by people who do not work for the company that made it. Its 2K rate is also cheaper than most of the field at that tier.
Take FastFrames when the loop is the product and the frame is not the deliverable — a live preview, a controllable world model, an avatar that has to answer inside a session, or a research programme that wants to change how a causal video stack decodes rather than what it generates. Accept 480p, silence, no reference inputs and a licence you have to read before you ship anything commercial.
FastFrames is not served by OrcaRouter, and this page makes no claim that it is — there is no hosted endpoint for Apple's weights. H3 is a different matter: we route it as minimax/minimax-h3, alongside the Kling 3.0 family and 200-plus language models, on one key at provider list price with 0% markup and automatic failover between endpoints. That is the cheap way to find out whether a per-second video model fits your pipeline before you spend a quarter's GPU budget on the alternative.

The one result that would make this a real contest
Two things would change the read, and neither is a better demo.
The first is FastFrames appearing on an arena board. That requires somebody to host it, and right now nobody does — Apple published weights and no endpoint, so there is no way for a preference test to serve it to voters. Until that happens, the model's quality claim rests entirely on its authors' VBench numbers, which is a weaker position than any model on H3's three boards occupies.
The second is MiniMax publishing a latency figure. H3's rate card is denominated in seconds of output, and nowhere in that structure is there a per-frame cost a reader could compare against 36 milliseconds. If MiniMax or an independent tester published frames per second at a given resolution and GPU, the two models would finally be measurable on a shared axis. They have not, and it is not in the vendor's commercial interest to start.

