Black Forest Labs' FLUX 3 Klein runs on an 8 GB consumer GPU
Black Forest Labs shipped the open-weights half of its FLUX 3 family, and Nvidia put a 120-billion-parameter agent model in the open. In research, a Goodfire essay argues the field misread its own evidence about what lives inside a network.
Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter image model that runs on consumer hardware — an INT8 build fits in 8 GB of VRAM and produces an image in under nine seconds on an RTX 4070. The company says Klein ranks first among open models on its own Elo testing, ahead of closed rivals Ideogram V4.5 and FLUX.2 [pro], and gains 91 Elo points on text-to-image plus 70 on editing over FLUX.2 [klein] with 38 percent fewer parameters. It handles text-to-image, editing, and composition from up to ten reference images, supports native 4K and pixel-accurate edits, and accepts bounding boxes for placing elements. We covered the API-only launch two weeks ago — Black Forest Labs launches Flux 3 Image with multi-step editing — and Klein is the open-weights release the company promised then.
The catch is the license, not the hardware. The weights ship under the FLUX Kommunity License, free for commercial use by companies earning under $5 million a year, which covers small studios and leaves agencies and mid-size shops paying BFL's API price, listed as starting at $0.009 per 1,000 images. As always with launch-day image models: the Elo comparisons are BFL's internal numbers, and the 8 GB and nine-second figures are the vendor's own timings, with no third-party replication yet. Our take: the interesting move here is the floor, not the ceiling — a frontier-class editing model now sits at the memory budget of a mid-range gaming card, and open weights are arriving weeks, not quarters, after the closed API launch.
Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter open multimodal model built for enterprise agents that run continuously in the background. It's a mixture-of-experts model with only 12 billion parameters active per token, and Nvidia says it matches its larger Nemotron 3 Ultra on agent tasks at roughly a fifth of the size. Two smaller models ship alongside: Bolt 74B-A9B, compressed to 2-bit so it fits in 24 GB of local memory, and a 4 GB Bolt E4B aimed at voice assistants, due October 23. The headline numbers are Nvidia's own — up to five times the throughput of similar-size open models, 80.8 percent on the CharXiv chart benchmark against 63 percent for Nemotron 3 Nano Omni, and, on 16 B200 GPUs running its Dynamo serving software, eight times as many concurrent users with 85 percent lower time to first token.
The sell is post-training on private data for buyers who can't ship documents to a hosted API, and Nvidia names Scale AI, Ramp, EY, Hebbia, CrowdStrike, and Perplexity among adopters. Treat the framing with care: Super VL still trails leading closed models by a wide margin on composite intelligence rankings, and Nvidia has not disclosed the task definitions or sample sizes behind its specialized-task wins.
A Goodfire researcher published a long essay arguing that polysemanticity — the canonical evidence that neural networks pack unrelated concepts arbitrarily — was a misreading. Nick Cammarata writes that the effect he and colleagues documented in InceptionV1 was mostly an ordinary learning failure in deep layers rather than superposition, and that ordinary architectural fixes make 99 to 100 percent of neurons in that layer individually interpretable, with no change to model size or training data. In preliminary audits of Qwen3-32B he estimates roughly 90 percent of neuron poles do one main job, about 7 percent do two related jobs, and only 2 to 4 percent remain unexplained, with similar results claimed in models up to a trillion parameters.
This is a working hypothesis, not a peer-reviewed finding. The companion paper is still promised, no third party has replicated any of the numbers, and the most-upvoted objection in the Hacker News thread is that the essay never defines what "clean" means. Why it matters: if networks are legible as built, interpretability stops being a fight to untangle weights and becomes ordinary science — which would redirect a large slice of safety research.
What to watch: whether anyone outside Black Forest Labs reproduces the Klein timings, and whether Nvidia's specialized-task numbers hold up when a third party writes the tests.
Weights are getting cheap and local. Does that make models easier to control, or just move the problem to whoever deploys them? Tell us in the comments.
