Tencent released Hy4 preview on August 28 — a 770B parameter open-weight model published under Apache 2.0 and available on Hugging Face right now. The model uses sparse Mixture-of-Experts routing: only 49B parameters activate per token, keeping inference costs well below what 770B total parameters would suggest. With a 1 million token context window and a built-in speculative decoding layer, it targets software engineering and agentic coding — and Tencent is offering free access on WorkBuddy and CodeBuddy for the next two weeks.
The competitive positioning is real. Hy4 scores 82.9 on SWE-bench Multilingual, ahead of DeepSeek V4 Pro (77.3), Kimi K3 (80.8), and GLM-5.3 (81.3). API pricing through OpenRouter starts at $0.83 per million input tokens — roughly one-eighth of what GPT-5.6 Sol charges per output token. Apache 2.0 licensing means no field-of-use restrictions and no geographic carve-outs, which is a genuine differentiator in this space.
How to Access Tencent Hy4 Now #
The fastest path is the free trial on WorkBuddy and CodeBuddy, Tencent’s developer products. Access runs two weeks from launch, meaning the window closes around September 11. No significant friction beyond creating an account — useful for quick evaluation before committing to API costs.
For API access, OpenRouter is the launch provider. Pricing: $0.834 per million input tokens, $2.501 per million output tokens. Cached input tokens cost $0.042 per million — a 20x discount that compounds quickly if you’re passing large system prompts or repeated context. Tencent Cloud TokenHub is also available as an alternative endpoint.
Self-hosting is technically possible under Apache 2.0 but infrastructure-intensive. The FP8 quantized variant requires approximately 720GB of VRAM — achievable on a single 8xH200 node, but beyond what 8xH100 nodes offer (640GB max). Prebuilt Docker images exist for both vLLM and SGLang. For most teams, OpenRouter is the practical starting point.
docker run --gpus all -p 8000:8000 --ipc=host \
vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 \
--tensor-parallel-size 8 --served-model-name hy4-preview
Tencent Hy4 Benchmarks: Where It Wins and Loses #
Hy4 leads open-weight peers where multi-language code matters. Its 82.9 on SWE-bench Multilingual beats DeepSeek V4 Pro (77.3), Kimi K3 (80.8), and GLM-5.3 (81.3). However, Terminal-Bench 2.1 tells a different story: Hy4 scores 85.4 while GLM-5.3 hits 88.2 and Kimi K3 reaches 88.3. For agentic coding tasks requiring sustained tool use, it’s competitive but not top-tier.
The generational jump from Hy3 is substantial. DeepSWE — a measure of real-world software engineering — climbed from 28.0 on Hy3 to 64.3 on Hy4, a 2.3x improvement in one model cycle. Claude Opus 5 still leads on SWE-bench Pro (79.2 vs Hy4’s 65.7), so if maximum coding accuracy is the requirement, the gap remains real. One important caveat: all Hy4 benchmark scores come from Tencent’s internal evaluators on proprietary task sets. No independent leaderboard verification exists yet.
The Preview Caveats Worth Knowing #
Tencent labels this an early version, and the limitations are real. Hy4 defaults to deep chain-of-thought reasoning and tends to over-verify its own work on straightforward tasks — inflating latency where none is needed. The fix is a single parameter change when calling the API:
response = client.chat.completions.create(
model="tencent/hy4-preview",
messages=[{"role": "user", "content": "Review this function for bugs"}],
extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}}
)
Additionally, OpenRouter measured roughly 86% availability in the model’s first three days — still stabilizing after launch. Tencent acknowledges headroom remains in both pre-training and post-training, meaning a more capable final version follows this preview. For production workloads with uptime requirements, wait for the stable release.
Key Takeaways #
- Hy4 preview is live now: free on WorkBuddy and CodeBuddy until around September 11, then $0.83 per million input tokens via OpenRouter
- It leads open-weight peers on SWE-bench Multilingual (82.9) but trails on Terminal-Bench 2.1 — strong for multi-language code review, middle tier for agentic tasks
- Apache 2.0 licensing is genuinely unrestricted — commercial use, modifications, and self-hosting all permitted without carve-outs
- Set
reasoning_effort: "no_think"
for direct responses; the default deep reasoning mode adds unnecessary latency on simple tasks - Self-hosting needs 8xH200 GPUs minimum for the FP8 variant — most teams will use OpenRouter until infrastructure scales