Week 32, 2026

Why: The longest single-run generation of any current video model at 4K, and the 50-reference input system is the most direct answer yet to character consistency, the problem that breaks most AI video work. Availability is the constraint: it ships inside ByteDance's own apps first, and the previous generation's international rollout was postponed indefinitely.

Why: If you want to know exactly what happened between the prompt and the image, this is the tool that shows you. Every step is a node you can open, change and re-run, and a workflow someone else built arrives as a file you can load rather than a screenshot you have to reverse engineer. That reproducibility is why it became the format the rest of the field builds around.

Why: It is the current number one on the independent Artificial Analysis Intelligence Index, and it got there while costing less per task than the model it displaced: $2.03 average per index task against Fable 5's $2.75. The effort dial is the reason to pick it over a fixed-tier model, because one integration covers cheap high-volume calls and expensive long-horizon agent runs.

Why: It is the shortest path from wanting a node workflow to having one running. ComfyUI asks you to bring a GPU and set it up; OpenArt hosts the compute and ships a template library, so the graph is something you edit rather than something you first have to stand up.

Why: Most canvases are built around one model family. This one treats the model as a node, so a pipeline can pass through several providers without leaving the graph. That matters when the best step for a job is not all from the same vendor.

Why: Its centre of gravity is the canvas rather than the graph, which suits the way a lot of real work happens: generate something, then keep editing regions of it. The node editor is there when a job needs repeating, instead of being the only way in.

Why: This is the release where Meta stopped giving models away. Muse Spark is closed, metered and sold through Meta's own API, and it lands at rank 15 on the independent index while undercutting comparable models on price. Worth tracking for that reason alone if your stack assumed Meta meant open weights.

Why: The architecture is optimised for throughput rather than peak benchmark score, and it shows: over 400 output tokens per second at 550B parameters. It is also the most openly documented release at this scale, because publishing the datasets and post-training recipes lets you actually reproduce and extend the model instead of just running it.

Why: The first credible attempt to collapse image, video and audio generation into a single model rather than a pipeline of separate ones, from the team behind the most widely self-hosted open image models. Access is the catch: video is gated early-access and the open-weight release has not shipped, so treat availability as limited until FLUX 3 Dev lands.

Why: The strongest open-weight coding model published to date: 62.1 on SWE-bench Pro against GPT-5.5's 58.6, and 81.0 on Terminal-Bench 2.1, at roughly a sixth of GPT-5.5's API price. The MIT licence carries no regional restrictions, so the weights can genuinely be self-hosted commercially, which is the reason to choose it over a closed model of similar strength.

Why: It is the first roughly trillion-parameter open-weights model that takes audio and images natively rather than through a bolted-on encoder, and Apache 2.0 means the weights can be used commercially without asking anyone. Thinking Machines is candid that this is not the strongest model available but a base worth customising, which is a more honest pitch than most open releases make.