Best open-source AI video and image models in 2026

Wan 2.2 T2V-A14B is the best all-around open-source pick for video: a genuine Apache 2.0 license, a fine-tuning base teams actually use, and a hosted option on Infer at $0.13/second if self-hosting isn't worth the GPU bill. SAM 3.1 leads segmentation, and it's the primitive under most "remove the background" and "isolate this object" pipelines running today. LTX-2 and its LTX-2.3 refresh post the highest Elo of anything open-weights on Infer's own leaderboard, but the license is revenue-gated rather than truly free past $10M ARR, and neither has a tryinfer.com catalog page yet. FLUX.2 rounds out the image side: Klein is Apache 2.0 and small enough to self-host on a single consumer GPU, Dev is bigger and better but needs a commercial license from Black Forest Labs before you ship it.

Ranked comparison table

If you're building for this, here's the model

• A fine-tuning base for a custom motion style, and you want to redistribute the result freely: Wan 2.2 T2V-A14B. Apache 2.0 carries no field-of-use restrictions, which none of the other three models here can say cleanly.
• The highest raw quality ceiling among open-weights video, and your company is under $10M ARR: LTX-2.3, still inside Lightricks' free commercial tier.
• Background removal or object isolation at scale: SAM 3.1, but budget the $0.01/image rate rather than the $0.002 figure until Infer reconciles the discrepancy.
• A small image model you can run on a single consumer GPU with no licensing call to make: FLUX.2 [klein], the only model on this page where "open source" and "Apache 2.0" mean the same thing without an asterisk.
• A defense, weapons, or export-controlled use case touching image or video segmentation: none of these four. SAM 3.1's license explicitly carves that out, and it's the kind of clause worth reading before a procurement conversation goes further.

Wan 2.2 T2V-A14B: the cleanest open-source video license

Wan 2.2 T2V-A14B is Alibaba's 14B-parameter text-to-video model, and its GitHub license file says exactly what it is: Apache 2.0, full stop, with no revenue threshold and no field-of-use carve-out. That makes it the one model on this list a legal team can clear in an afternoon. Infer positions it as a fine-tuning base for cost-sensitive prototyping and on-prem deployment, and it's hosted at $0.13/second if the self-hosting math doesn't pencil out for a given team.

The honest flaw is that Wan 2.2 isn't the frontier of its own family anymore. Alibaba has already shipped Wan 2.6 and moved Wan 2.7 into rollout, and it's Wan 2.6 that shows up on Infer's own leaderboard (rank 10, 1189 Elo), not the 2.2 version this page covers. Infer's own comparative copy puts Wan 2.2 at roughly 85% of Hailuo 02 Pro's quality, so the honest read is: cleanest license on this page, not the sharpest output.

Run Wan 2.2 T2V-A14B on Infer →

LTX-2 / LTX-2.3: the highest ceiling, gated by revenue

LTX-2 was released as, in Lightricks' own words, the first production-ready model combining open audio and video generation with native 4K output. Its Elo backs that up: on Infer's own 2026-04-30 snapshot, LTX-2 Pro (1129) and LTX-2.3 Fast (1121) rank 11th and 12th, ahead of nothing else open-weights on this page but genuinely close to Veo 3.1's 1208.

"Truly open" undersells the fine print, though. The release terms read: that's a revenue-gated license, not the no-strings Apache 2.0 grant the "open weights" framing implies for a company already past that threshold. LTX-2.3 is featured on Infer's homepage grid, but there's no catalog page to link, and so no CTA to give.

SAM 3.1: the segmentation primitive, with a license Infer mislabels

SAM 3.1 adds Object Multiplex to Meta's segmentation line: one prompt call segments many distinct subjects instead of one call per object, and it does it at roughly 0.4 seconds per frame. Three input modes ship today (click point, bounding box, coarse mask); text prompting is still on the roadmap, not live. That combination makes it the backbone of a lot of e-commerce background-removal and dataset-labeling pipelines that never mention Meta by name.

Infer's own model page calls the license Apache 2.0. The actual license Meta ships with the SAM 3 repository is not that: it's a custom SAM License, and it says exactly the kind of field-of-use restriction Apache 2.0 doesn't carry. Commercial use is broadly permitted under the same license, so most teams won't hit the carve-out, but "Apache 2.0" and "the SAM License" aren't interchangeable claims, and a team relying on Infer's copy alone would get the wrong answer if asked. Add the pricing page's own $0.002-vs-$0.01 discrepancy and SAM 3.1 is the model on this list most worth reading the source documents for, not just the marketing page.

Try SAM 3.1 in the Infer playground →

FLUX.2 [dev] and [klein]: two licenses, one model family

Black Forest Labs shipped FLUX.2 as a family, not a single release, and the two open-weight tiers land on opposite sides of the licensing line. FLUX.2 [klein] is described on BFL's own release post as a genuinely open-source Apache 2.0 model, size-distilled down to a scale that runs on a single consumer GPU. FLUX.2 [dev], the larger 32B model, ships its weights openly on Hugging Face but gates commercial use behind BFL's own licensing desk: the post says, in effect, download it freely and talk to BFL before shipping it in a product.

Neither tier is in Infer's catalog. Infer's live image lineup still runs FLUX 1.1 [pro] and FLUX.1 Kontext [pro], a generation behind FLUX.2, so there's no hosted price or CTA to give for either Dev or Klein; this is a self-host-or-license-direct pair, not an Infer-hosted one.

License table: what these licenses actually permit

"Open source" gets used loosely across this category. Here's what each license on this page actually grants, not what the marketing copy implies:

The self-hosting reality

Wan 2.2 is the only model here where self-hosting is a live financial question rather than a licensing one, and the math is unforgiving below a certain volume. A 720p render needs datacenter-class VRAM: an H100 PCIe at roughly $2.01/hour on-demand carries a tight enough margin that FP8 quantization is close to mandatory, while an H200 SXM at about $4.54/hour runs the same 720p, 10-second clips with real headroom. At the H100 rate, that's roughly the cost of 15 seconds of Infer's own $0.13/second hosted rate for every hour the card is rented, whether it's actively rendering or sitting between jobs. A queue that keeps the GPU busy for most of each rented hour is where owning the hardware starts to beat the API; anything sparser, and the hosted rate wins even before counting the engineering time to keep a GPU fleet healthy.

SAM 3.1's self-hosting case is milder since segmentation is far lighter than video diffusion, but the same logic applies in miniature: at $0.01/image, a self-hosted GPU only pulls ahead once labeling volume is high enough to keep it saturated, and idle GPU time has no equivalent to Infer's completed-generations-only billing.

How we ranked

Rank here weighs three things: how permissive the actual license is (not the marketing description of it), whether the model is hosted anywhere with a real price, and where a documented benchmark exists (Artificial Analysis Video Arena via Infer's leaderboard, 2026-04-30 snapshot). Where no confirmed number exists, we say so rather than inventing one. Infer hosts Wan 2.2 T2V-A14B and SAM 3.1, and neither one earns Infer a bigger cut than the other, so there's no incentive baked into this order; the license discrepancies flagged above came out of reading the actual license files, not out of favoring one provider's story over another's.

What didn't make the list

Seedream 4.0 and Nano Banana 2 are excluded on purpose: both are strong image models on Infer's catalog, but neither ships open weights, so they don't belong on an open-source ranking regardless of quality. HappyHorse-1.0/1.1 tops Infer's raw video Elo table at 1368, ahead of every model on this page, but it has no public API and no confirmed open-weights release, so there's nothing to self-host or license. Kling 3.0 Pro and Veo 3.1 remain the quality ceiling for video overall; they're proprietary, and they're covered on our full best AI video models ranking instead.

Related reading

• Best AI video models in 2026: the full proprietary-plus-open ranking, with Seedance 2.0 Pro at the top.
• Best cheap AI video models: where Wan 2.2's hosted rate lands on a quality-per-dollar basis.
• Best AI models for anime and stylized video: Wan 2.2 as a fine-tuning base for stylized work.
• Best AI image-editing models: where FLUX's proprietary [pro] and Kontext tiers rank against SAM's segmentation primitive.
• Best text-to-image models: the proprietary side of the image comparison this page's FLUX.2 section sits next to.
• All AI models directory: every ranked list on the site.