
FIBO 2.0 vs MiniMax H3 Image: the Crown Already Moved
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 514 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 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
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 116 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 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 · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
For twenty-one days, FIBO 2.0 was the best image-editing model you could download. Bria released it on 20 October 2026 at 1,086 Elo on Artificial Analysis's editing board, which put it first among open-weights entries and about 169 points above the company's own previous attempt. Then MiniMax put the weights for MiniMax H3 Image on Hugging Face on 10 November at 14:50 UTC, and by the time the arena had enough battles to rank it, the open-weights lead belonged to a model that had not existed when FIBO launched. Read on 20 November 2026, H3-Image sits at 1,135 Elo with an interval of 1,119 to 1,151. FIBO 2.0 sits at 1,086, interval 1,073 to 1,099. Those ranges do not touch.
That is a genuinely awkward position for a launch post to be in, and the useful question is what is left of FIBO 2.0's case now that the number it led with has been beaten by a competitor three weeks younger. The answer is not nothing. It is that the two models are built for different owners, and the difference shows up in a single figure: 8.1 billion parameters against a 68.65-gigabyte download.
What 8B and 68.65 GB mean for a person
Bria publishes the architecture in unusual detail. FIBO 2.0 is a flow-matching model running in the latent space of the FLUX.2 VAE, with Qwen3-VL-4B-Instruct as its prompt encoder, a two-layer Perceiver resampler compressing that encoder's output into 256 to 512 "gist" tokens, and a 44-block single-stream diffusion transformer with five blocks additionally fed by cross-attention. The checkpoint is 8,127,342,213 parameters in BF16 safetensors — roughly 16 GB of weights before any runtime overhead. A 24 GB consumer card holds it with room for the text encoder.
MiniMax H3 Image is a different order of thing. Its repository carries 68.65 GB of weights plus a text encoder borrowed from a second repository, which is what a model designed around 4K output, sixteen million pixels and up to sixteen reference images costs to ship. The 10 November announcement led with ten seconds to a sixteen-megapixel image, and whatever that number turns out to be worth on your hardware, it is not a number a single-card workstation produces.
• Weights on disk — FIBO 2.0 about 16 GB BF16; H3-Image 68.65 GB plus a separate text encoder.
• References per edit — FIBO 2.0 up to 4, addressed in order; H3-Image up to 16, per MiniMax's announcement.
• Output size — FIBO 2.0 renders at 1K-class resolutions through its published pipeline; H3-Image advertises 4K, roughly sixteen megapixels.
• Editing Elo — FIBO 2.0 1,086, 31st overall; H3-Image 1,135, 14th. H3-Image Highspeed, the faster tier, sits at 1,132.
• Text-to-image Elo — FIBO 2.0 1,037, 32nd; H3-Image 1,066, 27th; H3-Image Highspeed 1,072.
• Listed price — FIBO 2.0 $32.00 per 1,000 images; H3-Image $26.54 on editing and $21.79 on text-to-image; H3-Image Highspeed $45.25 on editing.
The reference count is the number that decides more real jobs than the Elo does. Four images cover person-plus-garment-plus-backdrop-plus-logo. Sixteen covers a whole product line in one call, and if your work is compositing rather than retouching, sixteen is not a nicety — it is the capability that makes multi-subject scenes possible at all. FIBO's four is enough for a fashion edit and not enough for a catalogue spread.

Three weeks is a long time on this board
The board FIBO 2.0 was reading in October has moved under it twice. H3-Image arrived on 10 November and went straight to fourteenth. Seedream 5.5, announced by ByteDance on 9 November, landed at 1,160 and eighth. Krea 3 (Ultra), five days older than FIBO, was already ahead at 1,154 with no weights at all. On the current read, the models above FIBO 2.0 on the editing board include three that shipped after it, and the open-weights entry immediately beneath it is Qwen-Image-2.1 at 1,082 — the model FIBO was being praised for passing a month ago.
None of that is a defect in the model. It is what a fast-moving category looks like, and the honest way to write about a three-week-old launch is to say so rather than to quote the launch-week snapshot as though the arrow had not moved. Bria's own Hugging Face card still carries the October claim — "first in image editing and second in text-to-image among open-weights models, October 20, 2026" — which was true on the date it was stamped. It is not true now, and the card does not say otherwise because model cards are not leaderboards.
What has not moved is Bria's other differentiator, and it is the one worth the most to a team that intends to keep using the model after this quarter's leaderboard is forgotten: the checkpoint is small, documented, and there is a fine-tuning path shipped with it. The repository includes a single-GPU LoRA trainer for editing and text-to-image with a data-preparation guide, on the argument that generation and editing share one set of weights, so one fine-tune can teach the model a house style and how to edit within it. A 68 GB model is not impossible to fine-tune, but it is a different budget conversation.
The licence neither vendor is shouting about
Both of these are open-weights releases in the technical sense and neither is open source. Bria's card states CC BY-NC 4.0 for non-commercial use, with commercial use requiring a separate agreement, and gates the repository behind a company-and-intended-use form. MiniMax H3 Image ships under its own community licence agreement — a bespoke document rather than an OSI licence — and the practical work of reading a bespoke licence is longer, not shorter, than reading a Creative Commons one. Neither company is offering you the thing the phrase "open weights" usually implies, which is the right to build a product on the file and ship it.

The distinction that does survive is cost of experimentation. A gated 16 GB download that fits on a workstation and comes with a fine-tuning script is something a single engineer can evaluate on a Tuesday. A 68 GB download that needs a multi-GPU box and a separate text encoder is a project with a ticket number. That asymmetry favours FIBO for research and prototyping, and favours H3-Image for anyone whose problem is already "we have the GPUs and we need sixteen references".
There is a third option that neither launch post mentions, and it is where most teams will actually end up: run neither locally. Both models are sold as hosted endpoints, at $32.00 and $26.54 per thousand images respectively, and at those rates the hardware argument evaporates for anything below a few thousand images a month. Self-hosting becomes rational when you need the fine-tune, the privacy, or the volume — not when you want the cheapest picture.
What a router does here, and what it does not
FIBO 2.0 is not on the OrcaRouter catalogue, and neither is H3-Image; the image models we do carry are a different set entirely. So the honest role for a router in this comparison is not "call both through one API", because that would not be true. It is the part around the render.
Both models want a structured, long-form prompt — FIBO through its VGL JSON schema, H3-Image through the long captions MiniMax's own examples use — and in production those captions are written by a language model, validated against a schema, and retried when the render comes back wrong. That loop is where the money and the failures are. OrcaRouter exposes more than 200 text models behind one key at provider list price with 0% markup, so the captioner and the validator can be swapped per job without a new contract, automatic failover covers a provider having a bad hour at 2 a.m., and the routing DSL composes the caption, the render call and the retry into one configured step. The image model stays where it lives either way.
That is a modest claim and it is deliberately modest. A router does not make a 1,135-Elo model into a 1,200-Elo one, and it does not get you sixteen references.
Who should pick which, this month
• Pick FIBO 2.0 if the machine under your desk is the deployment target, if you want to fine-tune on your own catalogue, or if "the licence has a named commercial path" matters more to your legal team than "the licence is short".
• Pick H3-Image if you composite more than four things into one frame, if 4K output is a requirement rather than a bonus, or if you have already accepted that the best open-weights model is a moving target and would rather buy the current leader's endpoint than its file.
• Pick neither if you edit a few hundred images a month and have no compliance reason to keep them in-house — both are sold as APIs at cents per image and the operational cost of a GPU box will exceed the invoice.

• Wait a fortnight if the decision is close, because three of the eight top editing models on the board shipped in November and the ordering has not settled.
What would change this read is concrete. If Bria relaxes the non-commercial term, FIBO 2.0's position changes overnight regardless of Elo, because the licence is currently the only thing standing between an 8B checkpoint and every product team that wants a model it can audit. If MiniMax publishes a quantised or distilled H3-Image that fits in 24 GB, the size argument disappears and FIBO's remaining case narrows to VGL and the fine-tuning scripts. And if the next arena read puts H3-Image's interval over FIBO's — they are 49 Elo apart with a 20-point margin between the bands, so it would take a genuine shift, not noise — then the October ranking stops being a historical curiosity and becomes a settled fact. For now it is a snapshot of a three-week window, which is exactly how the piece should be read.
