What Is Tencent Hy4 Preview? 770B Open MoE at $0.83
The Short Answer
Tencent Hy4 preview is an open-weight large language model that Tencent Hunyuan released and open-sourced on August 28, 2026. It is a mixture-of-experts model with 770 billion total parameters and 49 billion active per token, carries a context window exceeding 1 million tokens, and is priced at roughly $0.834 per million input tokens and $2.501 per million output tokens.
Tencent’s own framing is that stronger pre-training plus a substantially larger post-training run produced “the largest generation-over-generation gain we’ve measured” — and puts Hy4 preview at the open-source frontier.
Key Facts
| Tencent Hy4 preview | |
|---|---|
| Released | August 28, 2026 |
| Vendor | Tencent (Hunyuan team) |
| Architecture | Mixture-of-experts, 770B total / 49B active |
| Active ratio | ~6.4% of parameters per token |
| Context | >1,000,000 tokens |
| API price | ~$0.834 in / ~$2.501 out per MTok |
| Weights | Published on Hugging Face (tencent/Hy4-preview) |
| Status | Preview |
Last verified: August 30, 2026.
Why the Parameter Split Matters
770B total, 49B active. That ratio is the whole story of how this model can be priced below a dollar per million input tokens.
Compute cost scales with active parameters — you pay to route each token through 49B weights. Memory cost scales with total parameters — you must hold all 770B somewhere. Sparse MoE lets a vendor sell frontier-adjacent quality at a fraction of dense-model inference cost, which is why nearly every large 2026 release uses it.
The practical consequence for you: the API is cheap, and self-hosting is not. Those are not in tension; they are two sides of the same architecture.
Pricing in Context
Reference task of 30,000 input tokens and 5,000 output tokens, priced against the open-weight frontier tier as of August 30, 2026:
| Model | Input / Output per MTok | Task cost |
|---|---|---|
| DeepSeek V4 Pro (off-peak) | $0.66 / $1.98 | $0.0297 |
| Tencent Hy4 preview | $0.834 / $2.501 | $0.038 |
| GLM-5.3 | $1.40 / $4.40 | $0.064 |
| Muse Spark 1.2 | $1.25 / $4.25 | $0.059 |
| Qwen3.8-Max | $2.00 / $6.00 | $0.090 |
| Kimi K3 | $3.00 / $15.00 | $0.165 |
For comparison, Claude Opus 5 runs $5/$25 ($0.275 per task) and GPT-5.6 Sol $4/$20 ($0.22). Hy4 preview lands at roughly one-seventh the cost of Sol on that reference task while claiming frontier-adjacent capability.
Treat the “frontier-adjacent” part as a vendor claim until independent evaluations land. Tencent has not published a full third-party-verified benchmark table for the preview.
The Expert-Data Angle
The most substantive differentiator Tencent names is not architectural. It is data sourcing.
Tencent says it partnered with internal domain experts — software engineers, game developers, finance analysts and security experts — and built training data around the work those people actually ship. That is a deliberate move away from scraped generic text toward task-shaped supervision.
Whether it produces measurably better real-world performance is exactly what the next few weeks of independent testing will show. But the direction is worth noting: in 2026 the differentiator among large open models is increasingly post-training data quality, not parameter count.
Agentic Research Orchestration
The capability Tencent highlights hardest is agent orchestration: Hy4 preview coordinated several Codex sessions in parallel, evaluated their results, and synthesised the output.
That is a supervisor role, not a coder role. It is a meaningfully different job from “write this function” — it requires holding multiple work streams in context, judging quality across them, and deciding what to keep. The >1M-token context window is what makes it feasible.
If you are building multi-agent systems, this is the part of the release to test first: a cheap model that is competent at managing expensive models is worth more than a cheap model that is merely competent at coding.
Self-Host or API?
The weights are public. The economics still favour the API for most teams.
At 770B total parameters, even 4-bit quantisation puts you well past a single 8-GPU node. You are looking at multi-node serving infrastructure with the operational complexity that implies. Against $0.834/$2.501 API pricing, the crossover point where owning hardware beats renting tokens sits far above what most teams sustain.
Self-host when you have a hard requirement the API cannot satisfy — data residency, air-gapping, regulatory constraint, or genuinely enormous sustained throughput. Otherwise use the endpoint.
Who Should Use It
Good fit: high-volume agentic workloads where per-token cost dominates; long-context document and codebase work that needs the 1M window; multi-agent orchestration where a cheap supervisor coordinates expensive workers; teams that need published weights for compliance or future portability.
Poor fit: anything where you need independently verified frontier accuracy today. It is a preview, the benchmark picture is thin, and “preview” in 2026 still means the serving stack and the model can both change under you.
The router pattern still applies. Send bulk volume to Hy4 preview, escalate the genuinely hard minority of tasks to Claude Opus 5 or GPT-5.6 Sol, and the blended cost lands far below running everything on the frontier tier.