Best Open Models

The best-overall picks in this collection are — Image: FLUX.2 [klein] 9B; Video: LTX-2.3; Text · LLM: GLM-5.1. Each group also lists runner picks for workloads where the leader isn't the right fit.

Every pick here trades some quality headroom against the closed frontier for control: downloadable weights, inspectability, fine-tuning, and freedom from a single provider. Where the closed tier is meaningfully stronger, the pick's own row says so.

Open is not one license. This list spans Apache 2.0 releases and restricted tiers like the FLUX dev non-commercial license. The chip on each row reflects the license we verified for that model — never assume rights carry across rows.

Running these models through Runware's managed API is covered by the platform terms, on Runware's own optimized compute billed by compute time. Self-hosting rights come from each model's own license — check the chip and the model page before deploying weights yourself.

Every pick carries a license chip sourced from this collection's license registry, each entry verified against the publisher's official release page on a recorded date. Entries we have not yet re-verified to a named license render the neutral Open weights chip. Re-check the model page before any external launch — license terms differ per model and per version, and this collection pins the open version of each family rather than auto-following newer releases.

Prices on the rows are read live from the Runware catalog, exactly as billed through the managed API. Open models here run on Runware's own optimized compute and bill on compute time, so cost scales with GPU seconds and configuration rather than fixed per-request provider rates.

Address models by their full identifier and pin versions in production. Open families iterate quickly, and a newer family release is not guaranteed to ship under the same license as the version on this list.

For teams whose first constraint is openness: compliance regimes that exclude closed models, air-gapped deployments, fine-tuning plans, or provider-independence policies. It curates the strongest open models available through Runware so you don't have to union the open picks across every other collection by hand.

Three rules. Weights must be publicly downloadable — open weight is the floor, open source is a per-model chip. Every member needs a license registry record with a source and verification date before its row ships. And each family gets one pick per job: sibling size variants are mentioned in the row copy instead of padding the list. A new model enters by filling an empty job or displacing an incumbent, so the list stays small by construction.

No 3D or vision picks yet: the open candidates in those categories are either too few or lack licensing we can verify, and we would rather state the gap than pad the list. Wan's open video family is served through partner infrastructure and joins once its license verification lands. The collection grows only when a new family earns entry.

Prices, statuses, and release dates on this page are read live from the catalog. Membership is editorial: it is re-audited whenever a model enters or leaves, licenses are re-verified on a quarterly pass, and the collection pins the open version of each family — if a family's next release goes closed, the open predecessor stays.