Fibo 2.0 Dev Reaches ModelScope: The Checkpoint Bria Actually Built for Fine-Tuning

The Fibo 2.0 Dev listing on ModelScope's China site was created at 17:03 UTC on 20 October 2026 — the same day Bria released the model — and a copy on the international site appeared two days later, at 09:55 UTC on 22 October. Neither mirror has moved far since: the China listing carried 691 downloads and the international one 59 when checked this week, and both carry a licence field that reads, unhelpfully, "other". The download counts are not the story. The story is what ModelScope's own engineering stack says about the checkpoints, because DiffSynth-Studio's Fibo documentation now states it plainly: only Fibo 2.0 Dev is trainable.

That is worth slowing down for. Bria (the company behind the model) shipped Fibo 2.0 on 20 October 2026 as an 8.13-billion-parameter open-weights model that does text-to-image generation and instruction editing with one set of weights, and published two checkpoints under that name — the full base and a distilled editor. ModelScope, the Chinese model hub whose DiffSynth-Studio project supplies the training scripts, has now integrated both and drawn the line developers were waiting to see: the distilled checkpoint is a deployment artifact, the non-distilled one is the thing you can teach.

What is actually on ModelScope, and what is not

Two listings exist, and they are mirrors of the same weights, not a re-build. The China site lists briaai/Fibo-2.0-dev under the text-to-image-synthesis task with the libraries pytorch, lora, safetensors and diffusers — the presence of lora in that list is a small tell about intent. The international site carries the same model ID with its own, much smaller, download counter. Both point back at the gated Hugging Face repository rather than hosting a separate copy, so anyone who needs the actual files still walks through Bria's non-commercial access form.

Two details are worth flagging because they will confuse someone reading the listing in isolation. The task tag says text-to-image only, which undersells the checkpoint: Fibo 2.0 Dev does generation and editing from the same weights, and the editing path is the one most teams care about. And the licence field says "other" rather than "CC BY-NC 4.0", which is a metadata gap, not a different grant — the grant is still Bria's non-commercial licence, and it is the single biggest constraint on anything you fine-tune on top of this checkpoint.

The timing also matters for how the news should be read. The weights went up on the Hub on 18–19 October, Artificial Analysis had a snapshot by 20 October, ComfyUI merged native support on 21 October, and ModelScope's international mirror filled in on 22 October. Fibo 2.0 Dev is a nine-day-old checkpoint in a distribution phase that is still moving, not an old release being revived.

Dev versus Edit Turbo is a real fork, not a naming quirk

Bria published two checkpoints under one release name, and the difference decides everything downstream:

• Sampling — Fibo 2.0 Dev runs 50 steps with classifier-free guidance at 5.0; Fibo 2.0 Edit Turbo runs 4 steps with guidance fixed at 1.0, which is to say no guidance at all.

• Task coverage — Dev does text-to-image and editing; Turbo does editing only, so a text-to-image prompt against Turbo is the wrong tool rather than a slower answer.

• Negative prompts — they work on Dev because CFG is on; on Turbo there is no CFG to steer, so a negative prompt has no lever to pull.

• Trainability — this is the fork that matters here. DiffSynth-Studio's Fibo page is explicit that only Fibo-2.0-dev is a training target; Turbo is described as not intended for fine-tuning.

• Cost per image at inference — Turbo is roughly an order of magnitude cheaper to run for the narrow thing it does, which is exactly why nobody should fine-tune it to chase that gap.

Underneath, they are the same shape: a flow-matching diffusion transformer operating in FLUX.2 VAE latent space, an 8,127,342,213-parameter BF16 DiT with 44 single-stream blocks, five of which take cross-attention into a Qwen3-VL-4B-Instruct text encoder whose output is squeezed into 256–512 "gist" tokens by a two-layer Perceiver resampler. The repository runs about 25.5 GB all in — a 16.3 GB transformer, an 8.9 GB text encoder and a 0.34 GB VAE — which is the number to budget for, not the 8B headline.

The training recipe: frozen encoders, a trainable DiT

ModelScope's contribution is not another mirror but a script set, and the shape of it is conventional in the way that makes it usable. Training reuses the same FiboPipeline used for inference and runs through examples/fibo/model_training/train.py. The VAE and the text encoder are frozen; the DiT is trained with flow matching. That choice is the sensible one — fine-tuning a Qwen3-VL-4B encoder alongside a 16 GB transformer on a single node is a different project — but it also means a fine-tune can redirect style, composition and edit behaviour without teaching the model a genuinely new language of prompts.

Two levels are supported, and both are DiT-only:

• Full fine-tune — pass --trainable_models "dit" and let Accelerate handle the distributed config; the whole transformer moves, nothing else does.

• LoRA — attach adapters to the DiT with --lora_base_model "dit", set target modules and rank, and optionally resume from a checkpoint; the provided LoRA scripts default to a single GPU.

Editing is a first-class training target rather than an afterthought. Editing runs inject edit_image and an optional mask through --extra_inputs, and a UnifiedDataset reads those columns from metadata; rows without masks are trained as ordinary editing samples instead of being dropped. Because generation and editing share weights, one fine-tune can teach a new domain and how to edit inside it — Bria's own model card makes that pitch for brand and style LoRAs, vertical specialists and narrow task editors, and ships a single-GPU LoRA trainer with a data-preparation guide in the repository's fine_tuning/ folder.

The two example datasets show the schema you are expected to produce: DiffSynth-Studio/diffsynth_example_dataset, with a fibo/Fibo-2.0-dev split carrying a metadata CSV of image and prompt, and a fibo/Fibo-2.0-dev-Edit split carrying metadata JSON of image, prompt, edit_image and optional mask. Nothing in that schema is exotic; the work is in producing the captions.

VRAM: the part the model card does not quantify

There is no published minimum. The DiffSynth documentation recommends low-VRAM configurations and automatic VRAM management against remaining memory instead of naming a floor, and the example configuration tells you what they expect: load and prepare in FP8 E4M3, compute in BF16, offload to disk, and cap total usage at total CUDA memory minus 0.5 GB. That is a recipe written by people who know the answer depends on your card, your resolution and whether gradient checkpointing is on — and it is honest about it. Plan for offload and accept that a fine-tune at one megapixel will be slow before it is good.

Structured JSON is the interface, so your dataset has to speak it

Fibo's control surface is not a prompt box. The model was trained on Visual GenAI Language (VGL) captions — structured JSON in which lighting, composition, style, camera and objects are independently addressable fields — and the inference pipeline returns the caption alongside the image it produced. The bria-vgl helper exposes create() for generation and modify() for edits, so you are not expected to hand-write the JSON; a hosted VLM does the expansion unless you author it yourself.

That has a direct consequence for fine-tuning that the launch coverage mostly skipped. If you train on free-text prompts, you are training the DiT against a caption format it was not built around, and the control you were hoping to gain is the first thing to go. Storing VGL captions in your metadata — either generated by the pipeline and then edited, or authored deliberately — keeps the addressable-field behaviour intact, and it is the difference between a LoRA that moves a style and a LoRA that moves a style without wrecking the rest of the frame. If your training set is images without captions, the captioning pass is the expensive step, and it is a language-model job.

Tooling landed faster than most new checkpoints get

For a nine-day-old model, the surrounding stack is unusually complete, which is part of why the ModelScope integration is a smaller deal than it looks and a more useful one than a bare mirror.

• diffusers — a BriaFibo2Pipeline went in as PR #14991, which is the path most inference code will take.

• ComfyUI — native support merged on 21 October as PR #17240 and shipped in v0.43.0, adding a fibo2 text-encoder type that loads the official Qwen3-VL-4B and a TextEncodeFibo2 node that handles VGL prompts and reference images. BF16 and INT8 repacks follow that loader layout, and GGUF architecture support exists for quantised local runs.

• ModelScope / DiffSynth-Studio — inference and training scripts in one place, which is the piece a team standardising on DiffSynth actually needs before it can plan a fine-tune.

The wall nobody mentions: CC BY-NC 4.0

Fibo 2.0 is open weights, not open source, and the gap between those two phrases is the whole commercial question. The checkpoints are CC BY-NC 4.0: non-commercial use only, with commercial deployment requiring a separate agreement with Bria, and the Hugging Face repositories gated behind a company and intended-use form. A brand LoRA trained on Fibo 2.0 Dev is a research artifact until that agreement exists — which is a strange place to put "exactly what developers want to build on", but it is the accurate place. Compare the open-weight alternatives that permit commercial use outright and the licence, not the Elo, becomes the deciding line for anyone shipping.

The quality picture is also narrower than the headline. On Artificial Analysis' AA-Image-Editing v2.1, Fibo 2.0 reads 1,088 Elo with a ±14 interval over 5,009 blind votes — twenty-third overall and first among open-weights editors, though it has slipped a place or two since the launch snapshot. Qwen-Image-2.1 sits at 1,081 ±13, and the intervals overlap almost entirely, so the open-weights lead is a tiebreak rather than proof. On AA-Image-T2I v2.1 the ordering flips: 1,039 ±12 for Fibo 2.0 against Qwen-Image-2.1's 1,042 ±12. Bria's own model card is less flattering still in the narrow categories — it lists text rendering and layout in the high teens to low twenties, and human anatomy eighteenth. A fine-tune can chase those weaknesses, but it cannot exceed what the encoder understands.

Where routing fits, and where it does not

The honest split: Fibo 2.0 itself is not in OrcaRouter's catalogue, so nothing here is a "call Fibo through us" pitch. What routing buys is the other half of the pipeline. The two language-model jobs around this checkpoint — writing VGL captions for a training set and expanding short prompts into structured JSON at inference — are ordinary text workloads, and they are the ones that decide whether a fine-tune is possible at all. OrcaRouter offers one API for 200+ models with 0% markup, so provider list prices pass through and a vendor cut reaches your bill the same day; automatic failover across providers means a captioning run over a few thousand images does not die on one upstream; and the routing DSL can pin a cheap model for bulk captioning while sending the hard cases to something stronger, with model fusion available when a single captioner keeps missing a field. The image model stays where it is; the text around it is routable.

What to watch next is Bria's own roadmap, which the model card lists as multi-turn local editing, background removal, a product-shot LoRA, mask erasing and outpainting — all of which would arrive on the Dev checkpoint first, because Dev is the one with a training loop. The ModelScope mirrors are the boring half of this story. The DiffSynth sentence "only Fibo-2.0-dev is trainable" is the useful half, and it is the reason to download the 25.5 GB repository rather than call the API.