What it does
Hy4 Flash is the cost-focused model in Tencent's Hy4 family, announced on 16 November 2026. It accepts text and images, generates text and has a 1M-token context window. The model card lists 220B total parameters, including 64B of n-gram embedding (Engram) parameters, with about 12B active per token. Tencent built it by distilling from Hy4 and recommends it for everyday productivity tasks, while pointing to the larger Hy4 for long software engineering jobs. It is sold through the Tencent Cloud TokenHub API, runs inside Tencent apps such as WorkBuddy and Yuanbao, and its weights are open under the Tencent Hy License. The flagship sibling, Hy4, has its own page on Ordalin.
Official source 1 ↗ · Official source 2 ↗Features
- Text and image input with a 1M-token context. Hy4 Flash natively accepts text and images and generates text. The context window is 1M tokens (1,048,576), the same as Hy4. Official source 1 ↗ · Official source 2 ↗
- Small active footprint. The model card lists 155.5B backbone parameters plus 63.5B Engram parameters, with 11.3B text and 0.43B vision parameters active per token. Tencent says the design cuts key-value cache size to one sixth of Hy4's. Official source 1 ↗ · Official source 2 ↗
- Tuned for speed and short outputs. Tencent reports decoding speeds past 150 tokens per second on its own serving stack, and says reinforcement learning cut average end-to-end output tokens by 50%. Official source 1 ↗ · Official source 2 ↗
- Open weights under the Tencent Hy License. The instruct model, an FP8 build and the pre-trained base model can be downloaded. Tencent recommends vLLM or SGLang for production serving and ships a fine-tuning pipeline. Official source 1 ↗
- Switchable reasoning. The reasoning_effort setting defaults to "high" with thinking enabled and can be set to "no_think" for direct responses. Official source 1 ↗
- Tencent's own cost comparison. On WorkIndex, Tencent's internal benchmark built from typical WorkBuddy tasks, Tencent reports a score of 42.1 against 39.7 for Hy4 preview, at 1/13 of Hy4 preview's average cost per task and 1/20 of Hy4's. Official source 1 ↗ · Official source 2 ↗
Pricing & free limits
Cached input is $0.003 per 1M tokens. 50% off via Tencent Cloud TokenHub through 30 November 2026; OpenRouter lists the discounted $0.035 / $0.1405. Official source 1 ↗ · Official source 2 ↗
Hy4 flash, Hy4 flash-FP8 and Hy4 flash-base under the Tencent Hy License. Model-as-a-service businesses above US$50 million in yearly revenue, or products with more than 100 million monthly active users, need a separate written licence from Tencent. Official source 1 ↗ · Official source 2 ↗
Billing details
The 50% API discount on Tencent Cloud TokenHub runs through 30 November 2026; the list price applies afterwards. Official source 1 ↗
Free access & limits
Free to use inside WorkBuddy through 30 November 2026. Tencent's announcement does not list a free API tier. Official source 1 ↗
Best for & limitations
High-volume everyday agent work: deep search, document processing and slide creation. Official source 1 ↗
Simple frontend demos and computer-use tasks where cost and turnaround matter more than peak capability. Official source 1 ↗
Reading images, screenshots and long documents inside a 1M-token context. Official source 1 ↗ · Official source 2 ↗
Self-hosting an open-weight multimodal model for internal use. Official source 1 ↗ · Official source 2 ↗
- Not a standard open-source licence. The model card states that Hy4 flash is released under the Tencent Hy License (licence field: other, tencent-hy-license). The licence text says it is derived from Apache 2.0 but is not the Apache License: it adds commercial terms for model-as-a-service businesses, which need a separate written licence above US$50 million in revenue over any 12 months or before use in a product with more than 100 million monthly active users. Internal use is exempt. Official source 1 ↗ · Official source 2 ↗
- Weaker than Hy4 on long coding work. Tencent itself recommends Hy4 for complex workloads such as long-horizon software engineering, and its WorkBuddy team reports a slight dip on long-horizon coding compared with Hy4. Official source 1 ↗
- Results are Tencent's own. WorkIndex is an internal Tencent benchmark and the blind test used 85 internal WorkBuddy users, so the quality and cost comparisons are not independent. Tencent also notes that the base model sometimes breaks the expected answer format in few-shot tests, which may distort those scores. Official source 1 ↗ · Official source 2 ↗
- Still a large model to self-host. Tencent's vLLM example splits the FP8 build across 8 GPUs, the GGUF build is offered for end-to-end testing only, and Transformers support is not in the main branch yet. Hy4 flash-base is a pre-trained base model, not a chat model. Official source 1 ↗
- No video or audio input, text output only. The model card describes text and image input and text output. The larger Hy4 lists video understanding; Hy4 Flash does not. Official source 1 ↗