Open Source Image Models: The License Matrix Everyone Gets Wrong (August 2026)
Last updated August 7, 2026
Licenses attach to checkpoints, not families. Truly commercial open weights: Flux 2 klein-4B, Z-Image, the Qwen-Image line through Dec 2025 (all Apache 2.0), and Krea K2 (permissive, verify). Flux 2 dev and klein-9B allow commercial outputs but not commercial model deployment; Ideogram 4's weights are gated non-commercial despite Apache 2.0 code; 'Wan 2.7 open source' is false — the Wan open line ends at 2.2. Everything runs hosted inside the invideo agent, where the weights question is moot.
Updated August 2026
Most of what gets written about open source image models is licensing fiction. The two most-publicized "open" releases of 2026 — Flux 2's dev checkpoint and Ideogram 4 — both restrict commercial use, in two different ways that coverage collapses into "it's open source," while the models you can genuinely build a business on — Flux 2 klein-4B, Z-Image, the pre-2026 Qwen-Image line — get described with the same word. This page is the corrected matrix: which weights you can download, what each license permits, and the distinction almost every article misses — the difference between outputs you can sell and a model you can deploy.
Which image models are genuinely open for commercial use?
Four lines, as of August 2026, ship weights you can self-host commercially with no paid agreement:
- Flux 2 klein-4B — Apache 2.0; the only unrestricted checkpoint in the Flux 2 family (klein-9B is not; see below).
- Z-Image (Turbo and base) — Apache 2.0, from Alibaba's Tongyi-MAI lab; 6B parameters, runs on a 16GB consumer GPU.
- The Qwen-Image open line through December 2025 — Qwen-Image, the 2512 refresh, every Edit release through 2511, and Layered, all Apache 2.0 on Hugging Face and ModelScope. (The 2026 Qwen generation — 2.0 and 2.0 Pro — is API-only.)
- Krea K2 Raw and K2 Turbo — released June 23, 2026 under what Krea's technical report describes as a permissive license; verify the model card before shipping commercial work. Krea's flagship v2 API models stay closed.
Honorable ancestor: FLUX.1 schnell (August 2024) was Apache 2.0 and started the pattern. Everything else marketed as "open" this cycle needs the fine print below.
The license matrix
| Model | Weights downloadable? | License | Self-host commercially? | Self-hosted outputs commercial? |
|---|---|---|---|---|
| Flux 2 klein-4B | Yes | Apache 2.0 | Yes | Yes |
| Flux 2 klein-9B | Yes | FLUX Non-Commercial | No — paid BFL license required | Yes |
| Flux 2 dev (32B) | Yes | FLUX Non-Commercial | No — paid BFL license required | Yes |
| Ideogram 4 | Yes (gated) | Weights non-commercial; inference code Apache 2.0 | No | No — no revenue threshold |
| Z-Image Turbo / base | Yes | Apache 2.0 | Yes | Yes |
| Qwen-Image 2512 / Edit-2511 / Layered | Yes | Apache 2.0 | Yes | Yes |
| Qwen-Image-2.0 / 2.0 Pro | No | Closed (API only) | — | Via paid API |
| Krea K2 Raw / Turbo | Yes | Permissive per Krea — verify model card | Likely — verify | Likely — verify |
| GPT Image, Imagen, Recraft, Reve, MAI, Luma, Krea v2 | No | Closed | — | Per platform terms |
Read the middle rows twice — two different restriction shapes hide there. Flux 2 dev lets you sell the images but not deploy the model: outputs are commercial, while serving the checkpoint in a commercial product requires a paid Black Forest Labs license. Ideogram 4 lets you do neither from self-hosted weights: the gated weights are non-commercial with no revenue floor — a $200 freelance logo is as out of bounds as an enterprise deployment — and commercial rights come only through Ideogram's paid subscriptions or API. Same headline ("open weights"), opposite consequences for anyone invoicing.
The three corrections that keep needing to be made
"Flux klein is Apache 2.0" is only half true. Klein ships in two sizes, and only the 4B is Apache 2.0; klein-9B sits under the same FLUX Non-Commercial License as the 32B dev checkpoint. Posts celebrating "klein is fully open" describe the smaller model and imply the pair.
"Ideogram 4 is Apache 2.0 open source" describes the code, not the model. The Ideogram 4 inference code is genuinely Apache 2.0; the weights are a separate gated download under non-commercial terms (Hugging Face model gate, August 2026). Quoting the Apache label without the weights caveat is the most repeated licensing error of the June 2026 news cycle.
"Wan 2.7 is open source" is false. Alibaba's Wan open-weight line ends at Wan 2.2 (Apache 2.0, registry-verified); Wan 2.5, 2.6, and 2.7 are API-only.
The pattern across all three: "open" gets attached to a family, then inherited by every member — but licenses attach to checkpoints, not families, and 2026's releases split licenses within families on purpose.
What does an open license actually buy you?
Three things a closed API cannot sell you at any price.
Fine-tuning and LoRAs. Black Forest Labs publishes LoRA training docs for klein — 10–200MB adapters, 1–3 hours on a consumer GPU, uploadable back to the BFL API for hosted inference — with undistilled "klein Base" checkpoints provided for training; Krea's K2 Raw exists for the same reason. The most developed example is Qwen's: Apache 2.0 weights at near-frontier quality produced the largest editing-LoRA ecosystem anywhere — one community author's camera-angle LoRA shows about 205,000 Hugging Face downloads — and Alibaba then merged popular community LoRAs directly into the Edit-2511 base, per the official model card. Community fine-tune → product → absorbed into the next release: a flywheel closed families structurally cannot run.
Data isolation. Self-hosted weights mean prompts, reference images, and outputs never leave your infrastructure — the deciding factor for regulated industries and unreleased-product work.
Cost structure. Self-hosting swaps per-image fees for hardware: Z-Image runs on a 16GB consumer card, Ideogram 4's nf4 build on a single 24GB GPU (research only, per its license), while Qwen's 20B checkpoints want 40GB+ in BF16 (quantized builds cover consumer cards). Open does not mean cheap; it means the bill is yours to shape.
A calibration note on quality claims: Qwen's GitHub README calls 2512 the "strongest open-source image model" across 10,000+ blind evaluations — on Alibaba's own arena. The independent data point is Z-Image's debut as #1 open-source model on Artificial Analysis (December 8, 2025); leaderboards have moved since.
Where do the closed models leave you?
For GPT Image, Imagen (shutting down August 17, 2026 regardless), Recraft, Reve, MAI Image, Luma, Krea's v2 API models, and Qwen 2.0/Pro, there are no weights at any price — commercial output rights flow from platform and API terms. Those terms are not uniform: Recraft's free plan assigns ownership of generated images to Recraft, with commercial rights only on paid plans, and in the MAI-Image-2 era commercial API access required an application to Microsoft. "Closed" doesn't mean "commercially simple" — it moves the license question from a weights file to a terms-of-service page.
License questions that keep coming up
Is Flux 2 open source?
Partially, checkpoint by checkpoint. Klein-4B is Apache 2.0 — fully open. Klein-9B and the 32B dev checkpoint have downloadable weights under the FLUX Non-Commercial License: outputs are commercially usable, but commercial deployment needs a paid BFL license. Pro, flex, and max are API-only.
Can I self-host Ideogram 4 for commercial work?
No. The weights are gated non-commercial with no revenue threshold, even though the inference code is Apache 2.0. Commercial use runs through Ideogram's paid subscriptions or API, both of which include commercial rights.
Is Qwen-Image really Apache 2.0?
The pre-2026 line is: Qwen-Image, 2512, every Edit release through 2511, and Layered are Apache 2.0 — commercial use, modification, and self-hosting included. Qwen-Image-2.0 and 2.0 Pro are API-only.
Is Wan 2.7 open source?
No. Alibaba's open-weight Wan releases stop at Wan 2.2 (Apache 2.0); everything from 2.5 onward, including 2.7, is API-only.
What's the difference between "open weights" and "open source"?
Open weights means the checkpoint is downloadable — under whatever license the vendor picked, which may forbid commercial use (Ideogram 4) or commercial deployment (Flux 2 dev). Open source in the sense people intend requires a permissive license like Apache 2.0 on the weights themselves. Check the weights license, not the repo badge.
Does any of this matter if I never self-host?
Mostly no — hosted use is governed by the terms of the API or platform serving the model, not the weights license. It starts mattering the day you want fine-tuning, data isolation, or your own GPUs.
The honest availability note
Every family in this matrix — Apache 2.0, restricted-weights, and fully closed alike — runs hosted inside the invideo agent, where the weights question is moot: you are using licensed hosted inference among 200+ models, with output rights flowing through the platform rather than a checkpoint license. That is the fair framing: the matrix decides nothing about commercial use through a platform — only what you can do when you take the weights home. Family depth lives on the linked pages and the AI models index; the AI image generator workflow is where all of them get put to work.
Version history: first published August 2026, covering the Flux 2 license split, Ideogram 4's code-vs-weights gate, the Qwen Apache line, Z-Image, Krea K2, and the Wan 2.7 correction.