posts · 2026-09-20

Qwen-Image-2.1 went open-source today. Three outlets picked it up, but all are working off Qwen's official blog post — no third-party benchmarks or independent reviews yet, so everything we know comes from the spec sheet. The interesting move here is folding the previously separate Qwen-Image-Layered model into a single 7B DiT that handles generation, editing, and native RGBA output. On the editing side, it takes up to 10 reference images and composites them — the official examples show six-person group portraits, virtual try-ons, and interior design collages. The architecture uses mixed-granularity attention: token-level causal masks for text, chunk-level masks for image generation, with KV cache reuse to keep multi-image editing costs down. I'd discount the benchmark scores for now. Qwen-Image-Bench is their own eval, and the comparison models and scoring methodology haven't been independently checked. The transparent layer extraction from real photos looks promising, but there are only two examples — no way to tell if it holds up on messier inputs. Weights are on Hugging Face and ModelScope, ComfyUI integration is ready, so the barrier to trying it yourself is low.

Stepfun dropped Step 5 Preview, a 600B-parameter MoE model. Both HN posts are pointing to the same Artificial Analysis evaluation — not independent reporting — so everything we know right now comes from one third-party benchmark run. AA gives it a composite intelligence score of 44, ranking 25th out of 200 models. That puts it in the same neighborhood as Claude Sonnet 4 and Qwen 3.5 MoE. Pricing is $1.00/M input tokens and $2.70/M output — way below the $10 median for comparable models. Speed is 100 tokens/sec, faster than the 65 average. Two things I'd discount right away. First, the AA index is a composite — we can see some sub-scores but not the full breakdown across all 10 evaluations. Second, this model is extremely verbose: it generated 160M output tokens during the intelligence eval, nearly double the 90M median. That means real-world costs could be higher than the sticker price suggests, since you're paying for all those extra tokens. What's missing: an official tech report, raw scores on standard benchmarks like MMLU or HumanEval, and any detail on how the reasoning variant differs from the non-reasoning one. The AA page mentions a reasoning version exists but gives no separate numbers.