
Hy4 vs Qwen4-Max: The Two Rows Hy4 Actually Wins
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 182 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 749 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 1902 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 149 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 109 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
On seven of the nine evaluations that make up the Artificial Analysis Intelligence Index v5.0.1, Qwen4-Max beats Tencent Hy4. It should: Alibaba's flagship scores 42.9 against Tencent's 34.6, and it carries a shared thinking-and-answer budget of 262,144 tokens per call. But two rows go the other way, and they are the two that tell you what each model was actually built to do. On FrontierCode-AA 2.0 — software engineering inside real open-source repositories — Tencent Hy4 scores 31.4% and Qwen4-Max 26.5%. On MLCR-AA, medical long-context reasoning, Hy4 scores 32.8% and Qwen4-Max 19.4%.
That is a strange pair of wins for a model that is 8.3 index points behind, and it is the reason this comparison is worth more than a scoreboard. Qwen4-Max is the stronger general model and the one you would pick for hard reasoning, knowledge work and anything with video in it. Tencent Hy4 is the cheaper model, the downloadable one, and — narrowly, on real-repository work — the better engineer. The decision between them turns on which of those you actually need, and on a pricing structure that makes the cheaper model cheaper than its token prices suggest.
Seven rows to Qwen, two to Hy4
Both sets of figures come from Artificial Analysis on 20 November 2026, same evaluation versions, same harness:
• Hard reasoning, HLE-Diamond — Qwen4-Max 44.9%, Hy4 32.7%. The widest gap on the board, 12.2 points.
• Physics, CritPt v2 hard subset — Qwen4-Max 43.0%, Hy4 39.0%.
• Agentic coding in a clean terminal, Terminal-Bench 4.0 — Qwen4-Max 48.5%, Hy4 37.4%.
• Agentic science in a terminal, Terminal-Bench-Science — Qwen4-Max 23.3%, Hy4 12.9%.
• SaaS workflow execution, AutomationBench-AA — Qwen4-Max 72.6%, Hy4 65.6%.
• Professional document reasoning, GDP.pdf all-pass — Qwen4-Max 26.2%, Hy4 21.6%.
• Agentic knowledge work, AA-Briefcase Elo — Qwen4-Max 1,686.1, Hy4 1,445.9. A 240-Elo gap, the largest in this series.
• Software engineering in real repositories, FrontierCode-AA 2.0 — Hy4 31.4%, Qwen4-Max 26.5%.
• Medical long-context reasoning, MLCR-AA — Hy4 32.8%, Qwen4-Max 19.4%.
The pattern in the two losses is worth naming precisely, because it is not noise. Qwen4-Max wins the terminal benchmarks, which are clean-environment coding tasks with a defined goal. Hy4 wins the repository benchmark, which involves navigating code nobody wrote for the evaluation. And it wins medical long context by 13 points, which is the one capability that maps onto a stated design goal: Tencent trained Hy4 on long documents and deep research, and says so in the model card. Alibaba's own materials for Qwen4-Max, meanwhile, lead on Work360, token-efficiency claims and a shared reasoning budget — a general-purpose flagship, optimised for breadth.

The gap between the price ratio and the cost ratio
Qwen4-Max is 1.8 times Hy4's input price and 1.8 times its output price. On measured cost per Intelligence Index task it is 3.2 times — $3.27 against $1.03. The extra multiplier is token efficiency, and it is the part of this comparison most people skip.
• Input — Qwen4-Max $1.50 per million tokens, Hy4 $0.834. Qwen is 80% more expensive.
• Output — Qwen4-Max $4.50, Hy4 $2.501. Same 80%.
• Cache reads — Qwen4-Max $0.075 at a 95% discount, Hy4 $0.042 at 95%. Qwen is 79% more expensive, and the discount rates are identical, so caching does not rescue the gap.
• Output tokens per index task — Qwen4-Max 104,940, Hy4 78,058. Qwen reasons for a third longer per task.
• Time per index task — Qwen4-Max 29.8 minutes, Hy4 14.0. More than twice as long.
• Output speed — Qwen4-Max 59 tokens per second, Hy4 93.
Multiply the two effects and the 1.8x price ratio becomes a 3.2x cost ratio. For a workload that runs ten thousand agent tasks a month, that is the difference between a line item you defend and a line item you renegotiate — and it holds even though Qwen4-Max is the more capable model, because capability you are not invoking is still billed.

There is one pricing detail on the Qwen side that cuts the other way and is easy to miss. Alibaba's metered card is region-dependent: $1.50 and $4.50 in Singapore, but $1.413 and $4.24 in Beijing, Hong Kong, Frankfurt and US Virginia, with cache reads at $0.071. If your traffic terminates in one of those regions the input gap narrows to about 69% rather than 80%. It does not change the direction of the answer.
Video in, weights out
The two models do not accept the same inputs, and this is the difference that most often settles a real procurement:
• Input modalities — Qwen4-Max takes text, image and video. Tencent Hy4 takes text and image. Hy4 has lost nothing on the vision side since 4 November — image understanding and computer use were the headline additions in that release — but it has no video path.
• Output — both text only.
• Context — Hy4 1,048,576 tokens; Qwen4-Max 983,616 by Artificial Analysis's count, though Alibaba's own documentation states 1,000,000. The difference is bookkeeping, not capability; both are million-token models.
• Weights — Hy4 downloadable since 18 November 2026 under the Tencent Hy License, a 774-billion-parameter checkpoint in BF16 with an FP8 build and a pre-trained base. Qwen4-Max has no published checkpoint and Alibaba's materials describe a proprietary flagship.
The weights line is the one with consequences beyond price. A downloadable model can be fine-tuned on your own traces, quantised to fit your hardware, audited, pinned to a version and run inside a network that has no outbound route. A proprietary one gives you an API contract and a retirement date. That is why the earlier Qwen4-Max coverage on this blog kept returning to the same asymmetry: open weights at 770 billion parameters remove the legal barrier, not the hardware barrier — but for a company that already owns the hardware, that is the whole difference.
Where a router changes the arithmetic
Tencent Hy4 is routable here as tencent/hy4 at $0.834 and $2.501 per million tokens with $0.042 cache reads — the provider's own list, passed through at 0% markup, so a Tencent price move reaches you the same day. Qwen4-Max is not in our catalogue; Alibaba serves it through its own Model Studio API and several third-party platforms, and we do not route it, so nothing in this article should be read as an offer to.
What one endpoint does buy you is the ability to stop treating the choice as permanent. The routing DSL can send work by shape: repository-level engineering and long-document analysis to tencent/hy4, hard reasoning and anything with video to Qwen4-Max through its own API key, with automatic failover between them. On the component scores above that split is not a compromise — it is the correct allocation, because the two models' weak evaluations barely overlap. Model fusion would go further still: a panel of the two would put Qwen's HLE-Diamond score and Hy4's repository score in front of the same answer.

Who should pick which
Take Qwen4-Max if the work is reasoning-heavy, if any of it involves video, or if the outputs are judged by people rather than tests — the 240-Elo AA-Briefcase gap is not a rounding error. Budget for 3.2 times the per-task cost of Hy4 and a model that takes twice as long to finish each task.
Take Tencent Hy4 if you want to own the weights, if the work is repository-scale code or long-document analysis, or if unit economics decide the procurement. It is the cheaper model on every measured axis, it is the one you can run yourself, and it wins the two evaluations that reward navigating material nobody prepared for it.
The open question is whether that second list grows. Hy4's weights have been public for two days; a fine-tune on a company's own repository traces would attack a gap of five points, which is a weekend of compute at current prices. Qwen4-Max cannot be attacked that way at all. That asymmetry — one model you can improve and one you can only wait on — is not in any benchmark table, and over the next year it will matter more than the 8.3 points.
Compared in this article1
Detected from this article
