Tencent Hy4 API now available on FastMetal

Tencent Hy4 (hy4) is live on FastMetal, in the API and in chat. Tencent released it on November 4, 2026, as the full version of the Hy4 preview from August, and it succeeds Hy3, which we also serve.

What Tencent says changed

• Finance. In a blind evaluation Tencent ran with 213 finance professionals over 579 tasks, Hy4 scored 3.04, against 2.97 for DeepSeek V4.1 Flash 1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3.
• Fewer made-up facts. On Tencent's internal tests, the hallucination rate fell from 12.95% on the preview to 7.2%, and errors from unsupported assumptions from 23.1% to 5.6%.
• Agent work. Tencent reports better instruction following, long-context comprehension and long-horizon agentic tasks, plus training on Blender and CAD software.

All of these are Tencent's own figures. It says weights and a system card are coming "in the coming days"; as of today the only downloadable Hy4 weights are the August preview.

What we verified on FastMetal

We serve Hy4 with a 1,000,000-token context. Before turning the route on we sent real requests for each capability:

• Tool calling works, including a forced tool_choice.
• Strict json_schema output is enforced, even with a schema unrelated to the prompt.
• It reasons before answering, and those reasoning tokens bill as output tokens.
• No image input. Tencent added vision in this release, but on our route an attached image never reached the model: a solid red square came back as "Yellow". The model is marked text-only, so pick a vision model from /models for images.

Pricing

Hy4 bills per million input and output tokens, with cached input at a lower rate. It costs more than Hy3. Current rates are on the Hy4 model page and billing is explained on /pricing. Usage is prepaid, so you can't be charged past your balance. The host hasn't published a policy on training with prompts, so the model page shows that as unknown.

Calling Hy4

It's the OpenAI-compatible Chat Completions API with "model": "hy4". Here it returns a structured answer:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy4",
    "messages": [{"role": "user", "content": "Write a short poem about autumn in Kyoto."}],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "poem",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "title": {"type": "string"},
            "lines": {"type": "array", "items": {"type": "string"}}
          },
          "required": ["title", "lines"],
          "additionalProperties": false
        }
      }
    }
  }'

Control how much it thinks with reasoning. On "What is 17*23?", "reasoning": {"effort": "low"} cut reasoning from about 100 tokens to about 35, and "reasoning": {"enabled": false} turns it off. If you omit max_tokens, FastMetal lets a reply run to 32,768 tokens.

To try it without code, pick Hy4 in chat.

Compare real answers first

The Hy4 model page shows its actual answers to FastMetal's 43 shared prompts (summarization, bug fixes, translation and more), next to Hy3 and every other model we serve. The full catalog is at /models.