AI Model Report

Infrastructure · NOVEMBER 16, 2026

Tencent’s Hy4 flash activates 12B parameters and cuts API rates

The 224B-parameter model targets routine work with lower token prices, but Tencent’s internal task-cost results still need validation on customer-facing workloads.

By Aiko Tanaka · Inference & serving · November 16, 2026

Tencent released Hy4 flash on November 16, 2026, with 224 billion total parameters, 12 billion active parameters and a context window of up to 1 million tokens. Its lower published API rates target routine work, but customer-facing quality still needs testing.

Tencent’s launch announcement lists approximately $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens. Those rates aren't the same measurement as Tencent’s claimed task-cost savings.

Architecture and serving limits

Tencent combines hybrid attention, LatentMoE and Engram, with 64 billion Engram parameters included in the total. It says the KV cache is one-sixth the size of Hy4’s.

Tencent reports: "Reinforcement learning cuts average end-to-end task output tokens by 50 percent." Tencent also claims decoding above 150 tokens per second using megakernel optimization, 4-bit quantization and DFly techniques.

That throughput doesn't establish completion time. Tencent’s ADP team cautions that actual task duration depends on input length, tool calls and runtime environment.

TokenHub’s model summary lists image and text input, text output, function calling, structured output and caching.

Vercel separately confirms availability through AI Gateway as tencent/hy4-flash, with reasoning that can be disabled for requests that do not need it.

Task costs, benchmark limits and discount deadlines

Tencent’s comparison is substantial but internally measured: the company puts Hy4 flash’s per-task cost at one-twentieth of Hy4’s.

The scope matters: Tencent’s strongest savings claims come from WorkIndex, its internal benchmark covering document, spreadsheet, presentation and other WorkBuddy tasks.

The Tencent Cloud ADP post gives WorkIndex scores of 42.1 for Hy4 flash, 39.7 for Hy4 Preview and 51.4 for Hy4. It adds a comparison caveat: a combined chart substitutes flash’s visual results for Preview’s.

Flash therefore scores above Preview but below Hy4 on Tencent’s measure. These results don't establish improved customer conversion.

"The operator takeaway: compare cost per usable customer-facing result, not cost per million tokens alone."

For a small-team owner or founder, a bounded pilot using historical prospect-research or inquiry-response tasks would test the operating-cost opportunity. Compare completion quality, retries, corrections, latency and total cost against the current process before routing live customer work.

Promotional spend also needs separating from list-price spend. Tencent’s ADP post says TokenHub’s half-price period will end November 30, 2026, at noon Beijing time. i10X also reports that cutoff.

Vercel’s deadline differs: it advertises its introductory discount through November 29. Its discount will end that day, before TokenHub’s, so pilot budgets should track both promotional and post-promotion costs.

Sources