Tencent's Hy Image 3.5 adds conversational editing at $0.024 an image
The hosted preview keeps edits in context across turns and takes up to five reference images — no open weights, and its Seedream-parity claim is unverified.
Copy markdown
Editing that remembers the last turn
The multi-turn mode applies each edit against the running conversation instead of a blank prompt, and takes up to five reference images for character or style consistency. For iterative asset work, you stop re-describing the whole scene every round.
$0.024 an image, returned synchronously
Calls go through Tencent Cloud's TokenHub API at roughly $0.024 per image (¥0.15 domestic), billed only on success, with a free trial through Oct 7. Images come back in the same request — no async polling to wire up.
The catch: hosted-only, and the benchmarks are Tencent's
Unlike the open-weights Hunyuan Image 3.0, 3.5 is API-only, so no local runs. Tencent claims about 30% over 3.0 and rough parity with ByteDance's Seedream 5.0 Pro, but those are internal designer evals and no independent board lists it yet.
Elsewhere: a fully open 7B that reasons like a 235B
Zhongguancun Academy's ZGCM-1 shipped weights, data and training logs under MIT — 7.39B params, 256K context, 97.1% on MATH-500 and 75% on AIME 2026. The authors put it level with Qwen3-235B and GLM-5.1 on math and agentic search, running on a single GPU.