Tencent Hy4 flash scores 2.79 out of 4 vs. 2.81 for Haiku 5.5 and costs 18 times less per task
Tencent has released Hy4 flash, a lightweight model you can connect through the Selectel AI router. In a blind test on 259 real-world tasks, users gave it 2.79 out of 4, almost the same as Claude Haiku 5.5 (2.81). But one task cost ¥0.26 instead of ¥4.73, which makes it 18 times cheaper. If your agent processes long documents and spends heavily on tokens, the model is worth testing on your own data. So far, the figures come only from Tencent.
AI-processed from Habr AI; edited by Hamidun News
Imagine an agent that reads hundreds of long documents every day: it parses contracts, looks up answers in a knowledge base, and assembles presentations. On a flagship model, the token bill quickly eats into the product's margin. On November 16, China's Tencent introduced Hy4 flash, the smaller model in the Hy4 family.
According to user ratings, it falls only slightly short of Anthropic's Claude Haiku 5.5 (2.79 out of 4 vs.
2.81), while one task costs ¥0.26 instead of ¥4.
73, roughly 18 times cheaper. The model is now available in the Selectel AI router, a single OpenAI-compatible API with pay-per-token billing. The "224B" in the name is the number of parameters (the model's internal settings), but only 12B are actually active for each word, which is why the model is fast.
You can test it on your own documents in a single evening.
What's inside
Hy4 flash has 224B parameters in total, but only 12B are active for each token (a word fragment). This is the "mixture of experts" principle: for each word, the model activates a small part of a large pool of narrow specialists. Tencent puts decoding speed at over 150 tokens per second. The model accepts text and images and works with a context of one million tokens.
Its main feature is Engram memory, which accounts for 64B parameters. The idea comes from the DeepSeek lab. It works something like a reference book: instead of "computing" knowledge from scratch every time, the model looks up ready-made rows in a table. Several rows are read for each token, and their addresses are known in advance, so the table can sit in ordinary RAM rather than in expensive accelerator memory. Tencent calls the overhead of this negligible. The compute part takes about 160B parameters.
For long contexts, the KV-cache matters too. It is the model's working "cheat sheet" of what it has already read, and at a million tokens it is what eats up memory. Thanks to hybrid attention and the LatentMoE architecture, Hy4 flash's cache is six times smaller than Hy4's and 24 times smaller than Hy3's. In addition, reinforcement learning cut the average number of output tokens per task in half: the model answers more briefly, and therefore more cheaply.
What the tests show
There are no independent measurements yet: the model is not in the Artificial Analysis index. All figures below are Tencent's data; the write-up was published on Atomic Agent. According to the company, Hy4 flash is overall slightly ahead of Hy4 Preview, which has 770B parameters and 49B active.
The most telling result is the blind test: 85 users of the WorkBuddy app rated 259 tasks across 27 categories on a scale of up to 4 points.
- Hy4 flash: 2.79 points, a task costs ¥0.26 (about 3.2 ₽).
- Claude Haiku 5.5: 2.81 points, ¥4.73 (about 58.4 ₽).
- GLM-5.3-Flash from Zhipu AI: 2.53 points.
- DeepSeek V4.1 Flash 1017: 3.09 points, ¥1.20 (about 14.8 ₽).
Hy4 flash nearly matched Haiku 5.5 and beat GLM-5.3-Flash. DeepSeek V4.1 Flash is ahead on quality, but a task on it costs roughly 4–5 times more than on Hy4 flash. The choice depends on what matters more to you: price or score.
Where the model is weaker and how to connect it
On long programming tasks, the gap to the larger model is noticeable: on Terminal-Bench 4.0, Hy4 flash scores 19.2, while the full Hy4 scores 44.4. That is why Tencent recommends flash for everyday work (search, documents, presentations, simple frontend, and computer use) and reserves complex development for Hy4.
There is also an access limitation: Tencent has not yet published the weights, so you cannot run the model on your own hardware, only through an API. For teams that need independence from an external provider, this is a drawback.
Connecting is simple. In the Selectel control panel, you create an API key for the AI router, copy the model alias, and send a request to the chat/completions endpoint with an Authorization: Bearer header. Because the API is OpenAI-compatible, it is enough to change the base URL, key, and model name in your existing client; the rest of the code stays the same.
What to do in practice
If you have an agent for routine long-context tasks, such as document parsing, knowledge-base search, or building presentations, run your own examples through Hy4 flash and compare quality and cost with your current model. For complex code and serious engineering agents, the flagship Hy4 is still noticeably stronger, and Tencent's figures should be treated with caution until independent measurements confirm them.
- Nov 18, 2026Five AI Models Tried to Fix the Same Bug: Two Succeeded, Three Broke More Than They Fixed
- Nov 18, 2026Developer gives a web AI chat hands: an MCP server with 14 tools and 21,000 lines of code
- Nov 18, 2026OpenAI releases GPT-6.2 Astra: second place in the rankings and a 20% lower price
- Nov 18, 2026Three "token-saving" utilities raised the AI agent bill by 18–46%
Key Questions Answered
What's inside?
Hy4 flash has 224B parameters in total, but only 12B are active for each token (a word fragment). This is the "mixture of experts" principle: for each word, the model activates a small part of a large pool of narrow specialists. Tencent puts decoding speed at over 150 tokens per second. The model accepts text and images and works with a context of one million tokens. Its main feature is Engram memory, which accounts for 64B parameters. The idea comes from the DeepSeek lab.
What the tests show?
There are no independent measurements yet: the model is not in the Artificial Analysis index. All figures below are Tencent's data; the write-up was published on Atomic Agent. According to the company, Hy4 flash is overall slightly ahead of Hy4 Preview, which has 770B parameters and 49B active. The most telling result is the blind test: 85 users of the WorkBuddy app rated 259 tasks across 27 categories on a scale of up to 4 points. Hy4 flash: 2.79 points, a task costs ¥0.26 (about 3.2 ₽).
Where the model is weaker and how to connect it?
On long programming tasks, the gap to the larger model is noticeable: on Terminal-Bench 4.0, Hy4 flash scores 19.2, while the full Hy4 scores 44.4. That is why Tencent recommends flash for everyday work (search, documents, presentations, simple frontend, and computer use) and reserves complex development for Hy4. There is also an access limitation: Tencent has not yet published the weights, so you cannot run the model on your own hardware, only through an API.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.