
FIBO 2.0 Is on Civitai Now: Read the Commercial-Use Line Twice
- 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
Fibo 2.0 went up on Civitai on 18 November 2026, and the announcement says you can use what you make commercially. Bria's own model card, which has been public since the weights shipped on 20 October 2026, says the checkpoints are CC BY-NC 4.0 — non-commercial — and that commercial use requires a separate agreement. Both statements can be true at the same time. Whether they are, for your job, is the only question this page is really about, and it is the question almost nobody asks before they render a client deliverable on a model they found in a gallery.
The distinction that resolves it is not subtle once you see it, but it is invisible from the model page. What Bria licensed is a pair of files. What Civitai operates is a service that runs those files on its own hardware and hands you an image. A licence on weights and a term of service on a generation platform are two different instruments, and the second one is not written by Bria. So "use what you make commercially" is a claim about the output of Civitai's generator, and "CC BY-NC 4.0" is a claim about the four-gigabyte things you download and run yourself. If you are confused about which one applies to you, you are probably about to download.
What actually landed on 18 November
Civitai's model page for Fibo 2.0 carries a publish timestamp of 2026-11-18T00:16Z and lists two versions, matching Bria's own checkpoint split. Dev is the full, non-distilled base — the one that generates and edits. Edit Turbo is the distilled sibling, edits only. Under them sits a separate Turbo LoRA, rank 270, that Comfy-Org repackaged so Dev can sample in a few steps instead of its usual fifty.
The files are not Bria's original uploads. Civitai's page says plainly that both versions are the Comfy-Org repackaging of Bria's weights, in bf16 plus an int8 build whose outputs differ slightly from the original. That is worth knowing if you are comparing your results against anyone else's: an int8 render is not bit-identical to the bf16 one, and neither is identical to the diffusers pipeline the model card documents.
• Dev — full base model, 8B parameters, bf16 or int8, text-to-image and editing. Bria's recommended settings are 50 steps at guidance 5.0.
• Edit Turbo — distilled from Dev, same architecture, 4 steps at guidance 1.0, editing only, and the card says it is not intended for fine-tuning. A negative prompt has no effect on it.
• Turbo LoRA — optional adapter for Dev, loaded on top of Dev or a Dev fine-tune, for few-step sampling. Neither Bria nor Comfy-Org publishes settings for it; the workflow authors are using the edit model's 4 steps at CFG 1 as a starting point.
• On-site generation — Civitai runs the model itself, which is what makes the commercial-use line possible at all. The page limits generation there to PG-13 content.
Five weeks separate the weights from the gallery listing, and the gap shows in the download counts. The official repackaging has a handful of reactions. The community uploads that appeared first have more: a 6-step "FinalCut" fine-tune with 775 downloads, a 4-step generate-and-edit workflow at 1,464, a native-2K workflow at 631, GGUF quantisations at 396. The people who wanted this model locally were already running it before the front door opened.
Two licences, two different objects
Here is the part to read twice, because the two texts point in opposite directions and only one of them binds you.
Bria's model card is unambiguous about the weights: CC BY-NC 4.0 for non-commercial use, commercial use by separate agreement, downloads gated behind a login and an accepted condition. That is a copyright licence granted by the rights holder over a file. It travels with the file. Take the checkpoint to your own GPU, fine-tune it, build a product on it, and the non-commercial term is what governs — unless you have signed something with Bria that says otherwise.
Civitai's listing also carries a permissions block, and it is set permissively: commercial use allowed, derivatives allowed, different licensing allowed. That field is the platform's standard metadata for how a model may be used on the platform, and it describes what the uploader authorised for images made there. It does not overwrite CC BY-NC 4.0 on the weights, because Civitai is not the rights holder and cannot grant what Bria has not. Reading it as a relicence is the mistake that gets a small studio into a letter.
So the practical rule is dull and reliable. Generating on Civitai and selling the output is governed by Civitai's terms, and you should read them rather than the promo line. Downloading the checkpoint and using it in a commercial product is governed by Bria's licence, and unless you hold an agreement, that use is not covered. If your plan is "download the int8 build, run it on our own box, ship the results", nothing about the 18 November listing changes your position by one word.

One more line in that description is easy to miss and matters for anyone building a product: the licence it links is described as covering "how it applies here". That is a hedge, not a grant. Treat it as the platform telling you to read further, because it is.
What the two checkpoints do
The architecture is the reason this model behaves differently from a prompt-and-pray diffusion stack, and it is worth thirty seconds because it explains the editing quality. Fibo 2.0 is an 8B-parameter flow-matching model that runs in the latent space of the FLUX.2 VAE. Prompts are encoded by Qwen3-VL-4B-Instruct, and a two-layer Perceiver resampler compresses that output into 256 to 512 "gist" tokens before a 44-block single-stream DiT sees them. Five of those blocks also receive VLM layers directly through cross-attention — the language model is not just a caption reader, it is wired into the denoiser.
What you get on top of that is a model that takes a short instruction and expands it into a structured VGL/JSON caption before drawing, then hands you the JSON back. That is the whole selling point: an attribute you did not mention stays where it was, because it is a field in a document rather than a word that may or may not survive tokenisation.
• Reference images — up to four, addressed in order as <image_1> through <image_4>, for try-on, product integration and style transfer.
• Masked edit — white mask areas are regenerated, everything else preserved.
• Output size — multiples of 16 at roughly one megapixel. Defaults are 1024×1024, 1424×800 and 1248×832.
• Style modes — Boring Reality, Kodak Film, Hard Flash, Hazy/Dreamy, Lens Distortion, each with its own prompt scaffolding and negative prompt.
The three "coming soon" items on the card are worth flagging rather than repeating as features: multi-turn local editing, background removal, a product-shot LoRA, mask-based erasing and outpainting. The card is inconsistent about multi-turn — one section lists it as coming, another implies it works — so assume you are chaining separate edits rather than holding a conversation with the model.
What the community built in 48 hours
Two days is not long, and the uploads that exist are mostly plumbing rather than art. That is the useful signal. A model that gets a GGUF quantisation, a LoRA trainer GUI and a workflow with a documented wall-clock time within 48 hours of listing is a model people intend to run, not one they intend to look at.
The number I would trust least and repeat anyway is from the native-2K workflow: 19 seconds per image, one reference, 2K and 4 megapixels, eight steps, on an RTX 6000 Pro. It is one person's machine, one prompt set, and an unstated batch size. It is also the only wall-clock figure anyone has published, and it tells you that this is not a model you need a cluster for.
The 6-step fine-tune is the more interesting artefact, because it implies the community already considers Dev's fifty-step default too slow for iteration and is trading quality for throughput without waiting for Bria to ship something. If you are evaluating Fibo 2.0 for a production loop, that trade is the one you will end up making too, and it is being explored in public rather than in a roadmap.
The board moved since October
Bria's card still claims first in image editing and second in text-to-image among open-weights models, as of 20 October 2026. On the editing board that holds. On text-to-image it no longer does, and the reason is not a decline in the model.
Artificial Analysis, read on 20 November 2026, has Fibo 2.0 at 1,036.98 Elo on text-to-image (95% interval 1,025.98 to 1,047.98), rank 32 of 116 models. Above it among models you can download sit Qwen-Image-2.1 at 1,046.71 and FLUX 3 Image Klein at 1,041.34. That is third, not second — Klein entered the board after Bria's claim was written, and a claim about a rank is only true until someone else ships.
• Text-to-image — Fibo 2.0 1,036.98; FLUX 3 Image Klein 1,041.34; Qwen-Image-2.1 1,046.71. Fourteen points cover all three, and every pair overlaps inside its interval.
• Editing — Fibo 2.0 1,085.98 (1,072.98 to 1,098.98), rank 31 of 86. Qwen-Image-2.1 is next among open weights at 1,082.17. The 3.8-point gap is well inside both intervals.
• Against its own predecessor — FIBO Edit 916.87 and FIBO 854.16. The +169 on editing is roughly thirteen times the interval half-width, so that part of the story is not noise.
• Price — Artificial Analysis lists $32 per 1,000 images, the same representative figure it has carried since the model first appeared.
So the honest reading of "top of the open-weights rankings" in the listing copy is: true for editing, formerly true for generation. Neither statement is a lie about the model's quality. Both are statements about a leaderboard, and leaderboards are a moving photograph.

The other thing a month of use has produced is a clearer picture of the price. Bria's own platform charges around $0.032 an image, and the card notes a second charge if you use the VGL helper to turn a plain instruction into structured JSON — a call to a vision-language model that is billed separately from the image. That second charge is the one people forget when they budget, because it happens before the image exists.
Calling it from a pipeline
OrcaRouter does not host Fibo 2.0, and nothing here should be read as saying otherwise. The routes to this model are Bria's own platform and licensing programme, the ComfyUI and diffusion ecosystem where the community builds live, and now Civitai's generator. What a router is for in a Fibo pipeline is the other half of the job — the text encoder is a vision-language model, the prompt expansion is a separate call, and the batch job around it is the part that fails at 3 a.m.
If you are building an image pipeline where Fibo is one generator among several, the useful property is that one key reaches 200-plus models at provider list price with no markup, so a vendor price change on any of them is live here the same day rather than at the next billing cycle. That does not make Fibo cheaper. It makes the comparison cheap, which is the thing you actually want when you are deciding whether a $32-per-1,000 image model earns its place next to a $9 one.
Where routing earns its keep specifically is failover. This is a model you may be running on your own GPU from a non-commercial licence, and a single-box deployment has one failure mode: the box. Putting the surrounding calls — captioning, moderation, the structured-prompt expansion, the downstream text work — behind an endpoint with automatic failover means the image job is the only thing you are babysitting.
Who this is for
Three kinds of reader, three different answers.
If you want to try Fibo 2.0 before committing to anything, the Civitai listing is now the cheapest path in the world: no gated download, no diffusers install, no GPU. Render a few dozen images, judge the editing behaviour for yourself, and ignore the commercial question until you have something you want to sell.
If you are shipping commercially and you want the weights, the 18 November listing is not your licence. Bria's agreement is, and the price of that conversation is the price of the model. That was true on 20 October and it is true now.
If you are shipping commercially and you only want images, Civitai's generator is the route whose terms you should actually read — carefully, and to the end, because the interesting clauses in a platform's terms are never the one in the announcement.

The model itself has not changed in four weeks, and it is a good one: an 8B open-weights editor that holds the top of the open editing board and sits inside a fourteen-point cluster on generation. What changed is who is allowed to hand it to you and on what paper. That is a smaller story than a launch and a more consequential one, which is usually how it goes with licences.
