Summary:

The new download route is distinct from Hy4 flash's earlier hosted API launch. Self-hosting brings framework and hardware decisions, plus a custom license rather than standard Apache 2.0.

Tencent's Hy4 flash weights are now publicly downloadable. On November 18, 2026, the official release repositories offered the instruction-tuned model, an FP8 version and a pre-trained base model. This is a new deployment route, distinct from the earlier hosted TokenHub API launch.

Which download should you start with?

The repositories were public and ungated when checked. The model takes text and images and produces text, with a documented context ceiling of 1,048,576 tokens. Tencent describes roughly 220 billion total parameters and about 12 billion active per token. Active parameters measure computation, not the amount of memory needed to store all the weights.

Six comparisons from Tencent's evaluation

The following figures are developer-reported, not independent leaderboard results. Bold marks the best number among these four models in each row. Higher scores are better; lower cost is better.

BenchmarkHy4 flashHy4DeepSeek-V4.1-Flash-1017GLM-5.3-Flash
DeepSWE v1.1 score64.674.973.867.8
Terminal-Bench 4.0 score19.244.440.035.4
APEX-Agents v1.1 pass@163.175.272.951.5
AutomationBench v1.0.6 score52.256.254.348.4
MMMU-Pro score75.878.676.777.0
DeepSWE cost per task (CNY; lower is better)0.2905.4091.3043.230

Tencent used each model's highest available reasoning setting. DeepSWE used mini-SWE-agent; Terminal-Bench used the Claude Code harness; APEX used its reference agent. The cost row is Tencent's API-price calculation in Chinese yuan, not a forecast of self-hosting expenses. Its assumptions include fixed pricing and a 7.2 CNY-per-USD conversion.

The cover shows Tencent's separate internal WorkIndex aggregate. Its methodology substitutes flash results for Hy4 preview on vision-required benchmarks when plotting figures, so the preview comparison is not entirely a direct measurement. Across the selected score rows above, Hy4 leads flash; the cheaper option does not win by capability alone.

Open weights do not mean a small installation

Tencent provides vLLM and SGLang deployment recipes; its FP8 examples use eight-way tensor parallelism. The Transformers example currently points to a specific pull-request build while main-branch support is pending. Check the current recipe, GPU compatibility and memory needs before committing infrastructure.

A useful pilot should measure accepted results, latency and total operating cost on your own workload. Download access gives you control over deployment, but serving, updates, monitoring and capacity planning become your responsibility.

Read the custom license before deployment

The release uses Tencent Hy License 1.0, not standard Apache 2.0. Its additional terms require a separate written license for a model-as-a-service operator whose worldwide gross revenue exceeds US$50 million in any consecutive 12-month period, or before use in a product or service with more than 100 million monthly active users. The definition of model-as-a-service has exclusions, and qualifying internal-only use has a separate exception. Review the full text against your business rather than treating “open-weight” as unrestricted permission.

Sources and image credit