Independent · Daily · Every regionLatest edition: 21 November 2026
AI WorldwireVerified AI news from every region Subscribe
Latest

Model Review

Tencent Hy4 Flash Review: Cheap Office Agent, Weaker Coder

Tencent's Hy4 flash costs about $0.07 per million input tokens, with weaker coding than Hy4. Review of strengths, limits, pricing, plus Alibaba's Qwen4-Max.

Two Chinese labs released models within a day of each other. Tencent's Hy4 flash competes on price, Alibaba's Qwen4-Max on coding and agent work. Almost every number below comes from the vendors.

Key takeaways

What Tencent released

Tencent announced Hy4 flash on 16 November as the cheapest model in its Hy family. Its architecture mixes hybrid attention, a LatentMoE design and "Engram" parameters, and Tencent says the KV cache is one twenty-fourth the size of Hy3's and one sixth of Hy4's [1]. Reinforcement learning, it adds, halves average output tokens per task, and decoding exceeds 150 tokens per second [1].

The tests Tencent chose reflect the target use. Its internal WorkIndex benchmark covers slide creation, document editing, spreadsheet repair and audio or video processing in its WorkBuddy assistant. On it, Tencent says Hy4 flash costs one thirteenth as much per task as Hy4 Preview and one eighth as much as DeepSeek-V4.1-Flash-1017 at peak-hour pricing [1]. In a blind test with 85 Tencent employees across 259 tasks, satisfaction was close to Claude Haiku 5.5's, according to Tencent [1].

What Alibaba released

Alibaba's Qwen4-Max is a mixture-of-experts model with 2.4 trillion backbone parameters, 71 billion active per token and an extra 328 billion N-gram embedding parameters [2]. According to Alibaba, it uses roughly 48% to 49% fewer tokens than Qwen3.8-Max on Terminal-Bench 4.0, and its API prices are lower [2]. QwenCloud lists $1.50 input, $4.50 output and $0.075 for implicitly cached input per million tokens [4]. LLM Stats reports a 1 million token context and an OpenRouter listing, and notes smaller Qwen4 models are promised as open source [5].

Strengths and limits

Hy4 flash Qwen4-Max
Price per 1M tokens (input / output) $0.07 / $0.281 [1] $1.50 / $4.50 [4]
Weights Promised, no date [1] Closed; smaller models promised [2][5]
DeepSWE v1.1 (vendor-run) 64.6 [3] 75.7 [5]
Terminal-Bench 4.0 (vendor-run) 19.2 [3] 48.5 [5]

Strengths. Hy4 flash is very cheap, fast and aimed at repetitive document work. LLM Stats puts the shared price of Claude Haiku 5.5, GPT-6 Luna and Muse Ember at $0.10 input and $0.50 output, so Hy4 flash undercuts them by roughly 30% and 44% [3]. Qwen4-Max scores 75.7 on DeepSWE v1.1 against 74.2 for Claude Opus 5.5, according to Alibaba [5].

Weaknesses. Hy4 flash scores 19.2 on Terminal-Bench 4.0 against 44.4 for Hy4 and 40.0 for DeepSeek-V4.1-Flash-1017, in Tencent's own comparison [3]. Tencent itself sends long-horizon software engineering to Hy4 [1]. Qwen4-Max trails Opus 5.5 on Terminal-Bench 4.0, 48.5 to 59.6, in Alibaba's table [5]. We found no independent evaluation of either model yet, and Qwen4-Max's comparators were not run on a common harness [5].

Pricing and availability

Hy4 flash runs on Tencent Cloud TokenHub and inside WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. Until 30 November it is free in WorkBuddy and 50% off on TokenHub, and an OpenRouter listing is "coming soon" [1]. Tencent says it plans to open-source the model but gives no date [1]. Qwen4-Max is on QwenCloud, Alibaba's apps and OpenRouter [2][5].

Best fit, and who should wait

Hy4 flash suits high-volume summarising, slide drafting and computer-use agents where a small error costs little. Teams that need long-horizon coding, open weights, or a track record of independent testing should wait or look at Hy4, DeepSeek or Qwen4-Max. Qwen4-Max fits coding agents where cost matters but closed weights are acceptable.

The global picture

The US Commerce Department's Bureau of Industry and Security issued an interim rule on 16 November, titled "Addressing Remote Access to Advanced Semiconductors". It sets licence requirements on advanced chips going to Indonesia, Malaysia, Singapore, Thailand and Brazil, and on data centres that remotely supply significant compute to firms headquartered in Country Group D:5 or Macau [7]. CBS News reported that the rule follows concern about remote access through Southeast Asian data centres [8]. We are not linking it to either release; the rule shows how compute access is now regulated alongside model prices.

In South Korea, the government said it would try to shut public Wi-Fi at test sites for Thursday's college entrance exam and require test takers to hand in all digital devices, after a tax accountant exam cheating case involving generative AI, the Korea Times reported [9]. In Singapore, a National University of Singapore team posted MMFlash, a method for faster vision-language decoding, as work in progress [6].

What to watch

Our take

In our analysis, Hy4 flash is a pricing move: it pushes the cost of everyday agent work below the common $0.10 and $0.50 tier, while Tencent itself routes hard coding elsewhere. Qwen4-Max is the stronger coder of the two on vendor numbers, but it needs independent testing before anyone relies on the claimed lead over Opus 5.5. Neither review changes the buying rule: run your own tasks first.

FAQ

How much does Tencent Hy4 flash cost? Tencent lists about US$0.070 per million input tokens, US$0.281 per million output tokens and US$0.003 for cached input, with 50% off on TokenHub until 30 November [1].

Is Hy4 flash open source? Not yet. Tencent says it plans to open-source it and gives no date [1].

Is Qwen4-Max open weights? No. Alibaba says smaller Qwen4 models will be open-sourced [5].

Are these benchmarks independent? No. The scores come from Tencent and Alibaba; we found no independent evaluation yet.

This is analysis, not investment or purchasing advice. AI Worldwire uses AI assistance in research and drafting; an editor reviews every article.

Sources

  1. Tencent Launches Hy4 flash to Power High-Speed, Cost-Efficient AI Deployment for Everyday Workflows, Tencent, 16 November 2026.
  2. Qwen4-Max, Qwen team, Alibaba, dated 15 November 2026.
  3. Hy4 flash launch: price, benchmarks and context, LLM Stats, 16 November 2026 (secondary; tabulates Tencent's self-reported figures).
  4. Qwen4-Max model page and pricing, QwenCloud, accessed 17 November 2026.
  5. Qwen4-Max launch: price, benchmarks and availability, LLM Stats, 16 November 2026 (secondary; tabulates Alibaba's self-reported figures).
  6. MMFlash: Faster Speculative Decoding for Vision-Language Models, arXiv, 13 November 2026.
  7. Addressing Remote Access to Advanced Semiconductors, interim final rule (RIN 0694-AK50), US Department of Commerce, Bureau of Industry and Security, pre-publication text, November 2026.
  8. After Nvidia scandal, Commerce Department moves to curb remote access to AI chips, CBS News, 16 November 2026.
  9. Gov't seeks to block public Wi-Fi at college entrance test sites to stop AI cheating, The Korea Times, 16 November 2026.

Disclosure: the AI system that assisted this article, Claude, is made by Anthropic, one of the companies covered here.

Subscribe to the AI Worldwire newsletter