Three Chinese labs now ship the open AI models everyone benchmarks against. Tencent's Hy4 arrived in full in November 2026, DeepSeek's V4 Pro set the independent coding record, and Alibaba's Qwen3.8 Max quietly became the most capable Qwen ever. Which one should you actually use?
We compared all three on verified benchmarks, real API pricing, and licences — and on one uncomfortable fact: Tencent's headline scores come from Tencent itself.
In This Article
- The three models at a glance
- Tencent Hy4: what it is
- Benchmarks: what the numbers actually say
- Pricing: what each model costs
- Licences and self-hosting
- Which model should you pick?
- How to try each one free
- Frequently asked questions
The three models at a glance
| Model | Size | Context | Licence | API price (per 1M tokens, in / out) |
|---|---|---|---|---|
| Tencent Hy4 | 770B total / 49B active | 1M tokens | Apache 2.0 (preview weights) | $0.834 / $2.50 list* |
| DeepSeek V4 Pro | 1.6T total / 49B active | 1M tokens | MIT | ~$0.40 blended (Flash) |
| Qwen3.8 Max | 2.4T total / 95B active | 1M tokens | Custom licence (Max weights) | $2.00 / $6.00 |
*TokenHub runs a 50% discount through 31 December 2026 — effectively $0.417 in / $1.251 out. Cached input is $0.042 per 1M.
Tencent Hy4: what it is
Hy4 is Tencent's flagship Hunyuan model and the successor to Hy Memory 2.0 and Hy3: a mixture-of-experts model with 770 billion total parameters, 49 billion active per token, and a 1-million-token context window. The preview launched on 28 August 2026; the full Hy4 API went live on 4 November 2026 on TokenHub and OpenRouter, adding vision and computer use — the ability to see your screen and operate your PC. We covered that launch in Tencent Hy4: 770B AI That Sees and Uses Your PC.
One caveat: the 770B / 49B / 1M specifications come from the August preview model card and were carried over from the preview. As of the full release, only the Hy4-preview weights are downloadable under Apache 2.0; the full-model weights are still pending.
Cheaper options in the same family: Hy4 Flash, launched 16 November 2026 at $0.07 in / $0.281 out per 1M tokens, and the predecessor Hy3 (295B total / 21B active, 256K context, released 6 July 2026) at $0.132 in / $0.528 out.
DeepSeek V4 Pro (August 2026) is the biggest model here at 1.6 trillion total parameters with the same 49 billion active per token as Hy4, under an MIT licence. Qwen3.8 Max is Alibaba's flagship: a 2.4-trillion-parameter MoE with 95 billion active per token, a 1M-token context, and reasoning on by default, generally available since 3 August 2026.
Benchmarks: what the numbers actually say
Tencent's published Hy4 preview scores are strong: Terminal-Bench 2.1 at 85.4 (which Tencent says ties Claude Opus 5 and beats DeepSeek V4 Pro), DeepSWE up from 28.0 on Hy3 to 64.3, SWE-bench Pro at 65.7, GPQA Diamond at 92.3, Toolathlon at 74.1, and APEX-Agents at 37.1 — second only to Kimi K3's 37.2. But these are vendor-published figures. As of October 2026, no Hy model had a published SWE-bench Verified entry on the independent leaderboard — the one benchmark the industry trusts most for coding.
DeepSeek V4 Pro, by contrast, holds the independent record: 96.4% on SWE-bench Verified, the highest score on the independent leaderboards. Its cheaper sibling, DeepSeek V4 Flash, scores 88.8% on the same benchmark. That is a verified, third-party result — not a vendor claim.
On neutral ground, LMArena's text leaderboard has all three essentially tied: Hy4 preview at 76±1 (ranked #26), DeepSeek V4.1 Flash at 76±1, and Qwen3.8 Max at 76±1. Hy3 sits lower at 72±1.
Tencent is candid about Hy4's limits: the preview was text-only, with no visual input, and it has an "over-verification tendency" that slows delivery. The full November release adds vision.
The honest summary: on independent coding benchmarks, DeepSeek V4 Pro is the proven leader. On Tencent's own tests, Hy4 looks best-in-class. On real-user preference, all three are tied. Treat any vendor benchmark as marketing until an independent leaderboard confirms it.
Pricing: what each model costs
Price is where the three models separate sharply — and where Hy4 wins on paper.
| Model | Input (per 1M) | Output (per 1M) | Cheapest way in |
|---|---|---|---|
| Hy4 (TokenHub list) | $0.834 | $2.501 | 50% off to 31 Dec 2026 |
| Hy4 Flash | $0.07 | $0.281 | Cheapest flagship-class API here |
| Hy3 | $0.132 | $0.528 | Budget coding and agents |
| DeepSeek V4 Flash | ~$0.40 blended | Best value for proven coding | |
| Qwen3.8 Max | $2.00 | $6.00 | Most expensive of the three |
At list price, Hy4 costs roughly a third of Qwen3.8 Max on input and under half on output — and with TokenHub's 50% discount through the end of 2026, the gap is wider still. Qwen3.8 Max is the outlier at $2 in / $6 out: roughly 28x the input price of Hy4 Flash.
Licences and self-hosting
Licence terms decide whether you can run a model on your own hardware and sell what it produces.
Hy4: the preview weights are Apache 2.0 — the most permissive open licence — so you can download, modify, and commercialise them freely. But the full-model weights are not yet released, so the most capable Hy4 currently runs only through Tencent's API partners.
DeepSeek V4 Pro: MIT licence — equally permissive, no strings attached, and the full model is available. For self-hosting, DeepSeek is currently the safest bet of the three.
Qwen3.8 Max: the flagship's weights carry a custom Qwen licence, not Apache 2.0. It is free for internal use and most commercial use, but products above 100M monthly active users or $20M monthly revenue must display the model name, and model-as-a-service or AI-assistant businesses above $50M trailing-twelve-month revenue need a separate licence. Smaller Qwen3 models (up to 235B) are Apache 2.0 with no conditions.
One more warning: Tencent's Hunyuan 3D, Video, and World community licences do not apply in the EU, UK, or South Korea. That does not affect the Hy4 language model, but it matters if you use Tencent's wider toolkit.
Which model should you pick?
For coding: DeepSeek V4 Pro. Its 96.4% SWE-bench Verified score is independently confirmed, it is MIT-licensed, and the Flash variant gives you 88.8% at around $0.40 blended. See our best AI coding assistants roundup for the tools to run it in.
For cheap bulk use: Hy4 Flash or Hy3. At $0.07 in / $0.281 out, Hy4 Flash is the cheapest capable API in this comparison, and Hy3 at $0.132 / $0.528 is the budget pick for agentic workloads that burn through tokens. Use the full Hy4 while the 50% TokenHub discount runs to 31 December 2026.
For self-hosting: DeepSeek V4 Pro today; Hy4 preview if you want Apache 2.0 and can live with preview weights. Check Qwen3.8 Max's custom licence thresholds before self-hosting it for a business.
For computer use and agents: Hy4 — the only one of the three built around vision plus computer use, with an APEX-Agents score of 37.1 (second only to Kimi K3). DeepSeek is the stronger pure coder; Hy4 is the stronger operator.
For the Qwen ecosystem: Qwen3.8 Max, if you already run on Alibaba Cloud or Qwen's tooling. It is the most expensive option here, so the case for it is integration, not price or benchmarks.
Our overall pick: DeepSeek V4 Pro for work that must be right, Hy4 for work that must be cheap or agentic. Try both free first and let your own workload decide.
How to try each one free
Hy4: 1M free tokens per model for new TokenHub accounts until 31 December 2026; preview weights free to download under Apache 2.0; also on OpenRouter as tencent/hy4. DeepSeek V4 Pro: DeepSeek's own API and OpenRouter, with MIT-licensed weights. Qwen3.8 Max: Alibaba Cloud Model Studio and OpenRouter (qwen/qwen3.8-max-0902).
Tip: run the same real task — a coding job, a long document summary, an agentic workflow — through all three during the free trials and compare the outputs. Vendor benchmarks tell you what a lab measures; your workload tells you what you actually get.
Frequently asked questions
Is Tencent Hy4 really better than DeepSeek?
On Tencent's own benchmarks, yes — Hy4's Terminal-Bench 2.1 score of 85.4 ties Claude Opus 5. On the independent SWE-bench Verified leaderboard, DeepSeek V4 Pro leads at 96.4% and no Hy model has a published entry yet. On LMArena, the two tie at 76±1.
Is Hy4 really free?
The preview weights are free to download under Apache 2.0, and TokenHub gives new accounts 1M free tokens per model until 31 December 2026. After that, the API costs $0.834 in / $2.501 out per 1M tokens at list price — half that while the 50% discount runs.
Can I use Hy4, DeepSeek, or Qwen in the UK or US?
Yes — all three are available through APIs (TokenHub, OpenRouter, Alibaba Cloud) to US and UK users. The licence restrictions that exclude the EU, UK, and South Korea apply to Tencent's Hunyuan 3D, Video, and World community licences, not to the Hy4 language model. Tencent's consumer app Yuanbao, by contrast, is China-only and needs a mainland phone number.
Which open model is best for self-hosting in 2027?
DeepSeek V4 Pro: MIT licence, full weights available, and the best verified coding score. Hy4 preview is the Apache 2.0 alternative, with full weights still to come. Qwen3.8 Max's custom licence adds revenue thresholds — check the terms before building a business on it.
Prices were checked in November 2026 and change often — confirm on the vendor's pricing page before subscribing. Hy4 benchmark figures are Tencent-published unless marked independent.

No comments:
Post a Comment