Open-weight model access and efficient multimodal agents

Tencent launches Hy4 flash with 12B active parameters, but open weights remain pending

The new 224B multimodal model promises cheaper agent workflows; Tencent has opened commercial API access, not a verifiable public weight release.

Open Model Weights published17 Nov 2026
Primary sourceTencent — Hy4 flash official launch announcement
Source published2026-11-16
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original dark blue futuristic vector composition with a central cyan processor and orbiting violet neural cells suggesting sparse model activation.
Original conceptual artwork of sparse inference and model access, not an official Tencent image or a verified benchmark. AI-generated editorial illustration; not a documentary photograph.

Tencent introduced Hy4 flash on November 16 as a more economical member of its Hy model family. The company describes a 224-billion-parameter system, including 64 billion Engram parameters, with just 12 billion active per generated token. It accepts multimodal inputs and has an advertised maximum context length of one million tokens. Sparse activation may reduce compute for individual tokens, but total parameters alone do not establish device memory requirements or make the model downloadable. Primary source ↗

The architecture combines hybrid attention, LatentMoE and Engram features. Tencent says the key-value cache is one-sixth the size of Hy4's, while optimizations including four-bit quantization and megakernel execution enable decoding beyond 150 tokens per second. It also reports shorter completions after reinforcement learning. These remain developer statements, not independent Open Model Weights throughput tests on a specified GPU, context length and workload.

For now, access is commercial and hosted. Tencent lists Tencent Cloud TokenHub and its own productivity products, with prices of roughly $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens. The developer also cites its WorkIndex benchmark and an internal user study to support task-cost comparisons. Those metrics depend on the selected tasks, prices and evaluation setup and should not be mistaken for independently established general model quality.

The announcement says Tencent plans to open-source Hy4 flash, but does not supply a downloadable checkpoint, a weight license or a concrete release date. Its official public Hub and GitHub presence do not establish an already available Hy4 flash artifact. For a source-first model registry, that difference is decisive: a usable API and a promised open-weight model are two separate release states. Our coverage therefore records a noteworthy announcement and a public-weight verification watch, not a new verified model entry.

OMW REGISTRY WATCH

What this changes for the evidence layer

Do not enter Hy4 flash into the verified open-weight registry until downloadable weight files, a specific license and a pinned source revision can be independently verified. Tencent's API prices and performance numbers describe a hosted product, not a reproducible local deployment.

PRIMARY SOURCE

Tencent — Hy4 flash official launch announcement

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗