While everyone in the West argues about slowing down frontier AI, Tencent quietly dropped something Indian builders should pay attention to. On September 19, the Chinese giant open-sourced its Hy4 AI model — a preview release that pairs a massive context window with surprisingly aggressive pricing.
01.What the Hy4 open-source model actually is
Hy4 preview is a mixture-of-experts style model with 77 billion total parameters but only 4.9 billion active per query — that's the trick that keeps inference cheap. The headline spec is the context window of more than 1 million tokens, which puts it in the same conversation as the flagship models from OpenAI and Google for long-document work. Tencent says it was trained on high-quality data built with experts from its software engineering, gaming, finance and security teams.
02.How it scored against the competition
Tencent ran an internal blind test with 163 experts across 203 engineering tasks, and Hy4 preview averaged 2.99 out of 4.00 — slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). Kimi K3 has been making its own waves lately — we covered its arrival on Amazon Bedrock — so Hy4 beating it, even in a Tencent-run evaluation, is worth noting. Just keep the obvious caveat in mind: the test was run by the model's own maker. The claimed strengths are software engineering, office productivity and financial analysis, game development (including generating playable prototypes from a single natural-language request), and scientific research.
03.The pricing is the real story
Here's the part that should make every Indian startup founder sit up. Hy4's API is priced at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits. Compare that to frontier models charging $25+ per million output tokens and you see the gap. On top of that, Hy4 is free to use on Tencent's WorkBuddy and CodeBuddy until September 30. Open-weight releases like this are becoming a serious pattern — we recently covered another open model with computer-use skills — and every release pushes the price floor lower.
04.It helped build itself
The detail that honestly surprised me most: Tencent says Hy4 participated in its own development for the first time, automating optimization of training methods, data strategies, evaluation frameworks and operators — an early recursive self-improvement loop. It also autonomously analyzed bottlenecks in its inference system and tuned operator fusion and communication, lifting end-to-end throughput by 31.8%. Take the self-reported numbers with salt, but the direction is clear: models are increasingly tools used to build better models.
05.What it means for Indian developers and startups
For a bootstrapped Indian startup, a capable open model with sub-dollar input pricing changes the maths of building AI products — you can serve real customers on Indian rupee budgets without burning venture money on API bills. The same week this launched, reports also surfaced of former Intel and AMD executives starting AI chip ventures in India, so the local ecosystem is building on multiple fronts. Two things to think about before you switch: a Chinese-hosted API means your data crosses borders, so for anything sensitive, self-hosting the open weights or using IndiaAI-backed alternatives like Sarvam or the BharatGen effort makes more sense. And cheap tokens only help if your prompting is efficient — the cost-optimization chapters in our GPT-6 Astra — Zero to Hero Guide teach exactly how to keep token bills low regardless of which model you run. And if you're planning to build on a model this cheap, our AI Agents & Automation Playbook has 10 ready-to-build blueprints with verified 2026 pricing for n8n, Make and the WhatsApp Business Platform.
Key Takeaways
- 77B total / 4.9B active parameters — MoE design keeps costs low
- 1M+ token context window — enough for entire codebases and book-length documents
- Blind test score of 2.99/4.00 beat GLM-5.3 and Kimi K3 (in Tencent's own evaluation)
- $0.834/M input, free on WorkBuddy and CodeBuddy till September 30
- Self-optimization loop reportedly improved throughput by 31.8%
Frequently Asked Questions
What is Tencent's Hy4 model?
Hy4 is a preview-release, open-source mixture-of-experts AI model from Tencent, released on September 19, 2026. It has 77 billion total parameters with only 4.9 billion active per query, which keeps inference costs low, and a context window of over 1 million tokens.
How much does the Tencent Hy4 API cost?
Tencent prices the Hy4 API at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits. It is also free to use on Tencent's WorkBuddy and CodeBuddy tools until September 30, 2026.
Where can Indian developers access the Hy4 model?
Hy4 is available through Tencent's WorkBuddy, CodeBuddy, Yuanbao and Tencent Cloud TokenHub, as well as on OpenRouter — so you can try it without any special setup beyond an account on those platforms.
How did Hy4 score against competing models?
In Tencent's internal blind test, 163 experts rated Hy4 across 203 engineering tasks and it averaged 2.99 out of 4.00 — slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). Keep in mind the evaluation was run by the model's own maker, so treat the ranking as directional rather than definitive.
What This Means for You
If you're a developer in India, this is a good week to experiment — spin up Hy4 on OpenRouter or grab the free access on Tencent's tools before September 30 and benchmark it on your own tasks, not Tencent's. If you're a founder, note that the price of intelligence keeps falling, and your moat will be what you build around the model, not the model itself. And whatever model you end up using, the skill that pays is knowing how to direct it — prompting, agents and cost discipline — which is exactly what the GPT-6 Astra — Zero to Hero Guide teaches, from first prompt to working AI agents.
And if GPT-6 Astra itself is still new to you, our GPT-6 Astra explainer is the best place to start — it covers what the model does, what it costs, and how to stay safe while using it.