Tencent Hy4-preview 770B: The Apache 2.0 MoE That Undercuts GPT-5.6 Sol by 4× in 2026

Tencent's Hunyuan team released Hy4-preview on Hugging Face under Apache 2.0 on August 29, 2026. The model card reports 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro with FP8 weights, shipping day-one Docker recipes for both vLLM and SGLang. At $0.834/M input and $2.501/M output tokens on Tencent Cloud TokenHub, Hy4 undercuts GPT-5.6 Sol by 4× on input costs.

This is the most aggressive open-weight frontier release since Qwen3.8-Max's $2/$6 pricing in August. Here is what it means for the AI industry.

Architecture: 49B Active of 770B Total

Hy4 uses a classic Mixture-of-Experts architecture with 256 routed experts plus 1 shared expert. Each token activates only 49B parameters — roughly 6.3% of the total — while accessing the full representational capacity of 770B parameters through expert routing.

Total parameters:  770B
Active per token:  49B (6.3%)
Routed experts:    256
Shared experts:    1
Context window:     1M tokens
Weight format:      FP8 (day-one release)
License:            Apache 2.0

The 256-expert routing is significantly wider than DeepSeek's 16-expert MoE or Qwen3.8's 64-expert configuration. More experts mean finer-grained specialization — Hy4 can route code-generation tokens to code-specialized experts and reasoning tokens to logic-expert shards without cross-contamination.

Benchmark Comparison

Hy4's 92.3 GPQA Diamond score edges out GPT-5.6 Sol's 91.8 — a remarkable result for an open-weight model. On SWE-bench Pro, Sol still leads at 68.2 versus Hy4's 65.7, suggesting proprietary post-training still gives an edge on complex code-generation tasks.

Cost Analysis: The Unit Economics

At $0.834/M input tokens, Hy4 costs:

• 71% less than GPT-5.6 Sol ($2.50/M)
• 72% less than Claude Opus 5 ($3.00/M)
• 17% less than DeepSeek V4-Pro ($1.00/M)

For a typical agent loop processing 50K input tokens and generating 5K output tokens:

Running 10,000 agent calls per day on Hy4 costs $546 — compared to $2,000 on GPT-5.6 Sol. That is a $45,000/month savings for a moderate-volume production agent.

Self-Hosted vs. Cloud Pricing

Tencent Cloud TokenHub pricing is competitive, but self-hosted deployment eliminates the per-token cost entirely:

• 8×H100 80GB cluster: ~$25/hour on spot pricing = $0.013/hour per GPU
• Hy4 FP8 inference throughput: ~12 tokens/second/generation on 8×H100
• Effective cost at 80% utilization: ~$0.08/M input tokens (vs. $0.834/M cloud)

Self-hosting cuts costs by 10× but requires managing GPU infrastructure, handling failures, and scaling capacity.

The Apache 2.0 Implications

Hy4's Apache 2.0 license is the most permissive possible — no attribution requirements, no commercial restrictions, no copyleft obligations. This enables:

1. Fine-tuning without restriction: Enterprises can fine-tune Hy4 on proprietary data and deploy the derivative models commercially.
2. Embedding in products: SaaS companies can embed Hy4 inference without licensing fees or revenue sharing.
3. Government and defense: Unlike some licenses with use-case restrictions, Apache 2.0 imposes no conditions on deployment domain.

Compare this to Meta Muse Glimmer 30B (Apache 2.0) and DeepSeek V4 (MIT) — the open-weight frontier is converging on fully permissive licensing.

Production Deployment Options

What This Means for the Open-Weight Race

Hy4's release accelerates the convergence between open-weight and proprietary frontier models:

1. GPQA Diamond parity: The first open-weight model to exceed GPT-5.6 Sol on a major academic benchmark.
2. 1M-token context: Matching Gemini 3.1 Pro's context window in an open-weight model.
3. 256-expert MoE: The widest expert routing in any publicly available model, enabling fine-grained specialization.
4. Day-one inference support: vLLM 0.28.0, SGLang, and Koboldcpp all support Hy4 on launch day.

The gap between open and closed frontier models is now measured in single percentage points on benchmarks and dollars per million tokens on pricing. For most production workloads, the open-weight option is now the default choice.

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last tested: August 2026 with Python 3.12, vLLM 0.28.0, and Hy4-preview FP8 weights on 8×H100-80GB.