Hy4 Preview Puts 1M-Token Agents on the Table

Hy4 preview, Tencent’s newly open-sourced flagship language model, puts a 770-billion-parameter mixture-of-experts model and a context window exceeding one million tokens into the hands of builders. The headline is not a reason to swap models overnight. It is a reason to revisit what an agent is allowed to keep in view—and what it must still prove before acting.

The big signal: Hy4 preview

Hy4 preview is an MoE model: Tencent lists 770 billion total parameters but 49 billion active parameters, with a context window above one million tokens. Tencent has released the weights and says the model is also available through its own products, Tencent Cloud TokenHub and OpenRouter. That combination is the signal. Big-context agent work is moving beyond a closed-model-only conversation.

One million tokens is easy to turn into marketing theatre. In a working system, it could mean more of a codebase, an operating manual, a customer history or a sequence of tool results can stay available without being repeatedly retrieved and summarised. It does not mean an agent suddenly understands a business. A longer window can just as easily preserve more stale instructions, duplicate evidence and unsafe tool output.

Tencent positions Hy4 preview around software engineering, office work and research. Its own blind comparison used 163 experts on 203 engineering tasks and put Hy4 slightly ahead of GLM-5.3 and Kimi K3. That is useful directional evidence, but it is still a vendor-run test on vendor-selected work. Teams should treat the result as an invitation to evaluate, not a universal leaderboard verdict.

The more concrete commercial detail is the listed API pricing: US$0.834 per million input tokens, US$2.501 per million output tokens and US$0.042 for cache hits. Cache pricing matters disproportionately to agents because they often resend a system policy, task plan and growing history. Still, a million-token context does not make every task economical; it can encourage expensive, slow prompts where a smaller model plus good retrieval would have done better.

Open-source watch

• Hy4 preview: Tencent’s launch announcement and GitHub repository make this the week’s major open-weight story. Its scale means “open” does not automatically mean “run it on a laptop”; serving it remains an infrastructure decision.
• GLM-5.3: our recent look at GLM-5.3’s open weights made the complementary point: model freedom needs controls around it. Hy4 adds a long-context option to that same question, rather than settling it.
• Qwen3.8 27B: smaller multimodal options remain important when the job is visual understanding close to the product. Qwen3.8 27B’s visual-agent trade-off is a useful counterweight to a giant text model: not every workflow benefits from a million-token window.
• Serving is part of the model choice: Tencent’s model distribution page is a reminder that weights, quantisation, hardware, batching and observability decide whether an open release becomes a dependable service. The checkpoint is only the first component.

What builders should take from this

For meLink, the interesting question is not whether an assistant can be given a million tokens. It is whether the assistant can distinguish durable customer-approved knowledge from transient conversation, retrieve the right source at the right moment and show its work when it recommends or acts. More context improves the ceiling; orchestration determines the floor.

That makes a tiered design more attractive than a one-model bet. Keep routine website questions, classification and extraction on a small, fast lane. Escalate bounded research, code analysis or multi-document work to a stronger long-context lane. Put approval gates around actions that change data, spend money or speak for a business. And record which lane produced the answer.

Before testing Hy4 preview, write a small evaluation pack from real work: a support conversation with conflicting facts, a product document that changed last week, a task that requires a human hand-off, and a tool call the agent must refuse. Measure groundedness, latency, cost, hand-off quality and recovery—not just whether a polished answer appeared.

There is also a practical preview warning. A preview model can change in behaviour, pricing or deployment guidance. Pin versions where possible, preserve a fallback route, and do not let a promising benchmark become a single point of failure in a customer journey.

The practical takeaway

Hy4 preview makes open-weight long-context agents a more serious option, especially for teams prepared to run a real evaluation and operate the serving layer. The useful takeaway is not “put everything in the prompt.” It is “keep the right evidence available, route work to the right capability tier, and make every consequential step reviewable.”

That is the adoption pattern worth watching: open models are no longer only a cost or sovereignty alternative. They are becoming another lever for designing agent systems that can be inspected, swapped and kept inside the boundaries a business actually needs.