Tencent Eval
Use the tencent-eval provider for Tencent's Hunyuan Hy4-flash evaluation Chat Completions API (hy4-flash-dev).
Direct model selection only
tencent-eval is intentionally omitted from the interactive model picker. Select it directly with --model tencent-eval.hy4-flash-dev, --model hy4-flash-dev, or --model tencent-eval.
Configure
Set either TENCENT_EVAL_API_KEY or HY4_API_KEY:
export TENCENT_EVAL_API_KEY="your-api-key"
# or Tencent's guide variable:
export HY4_API_KEY="your-api-key"
Or configure it in fast-agent.secrets.yaml / fast-agent.yaml (config file credentials take precedence over TENCENT_EVAL_API_KEY, which takes precedence over HY4_API_KEY):
Optional provider settings:
tencent-eval:
api_key: "${TENCENT_EVAL_API_KEY}"
base_url: "https://eval.hunyuan.tencent.com/openapi/v2/data"
default_model: "hy4-flash-dev"
# default_headers:
# X-Custom-Header: value
You can also override the endpoint via TENCENT_EVAL_BASE_URL when tencent-eval.base_url is not set in config.
Models
| Model | Context | Maximum output | Input | Structured output |
|---|---|---|---|---|
hy4-flash-dev |
960,000 | 65,536 | Text, images (PNG verified) | JSON Schema (strict: true) & JSON Object |
The 960,000-token context limit is server-verified from endpoint input validation (Input tokens exceed the configured limit of 960000 tokens). fast-agent omits max_tokens on the wire by default unless explicitly set (verified up to 35,315 completion tokens on a single turn without max_tokens), and GET /models returns 404.
Select the model with a provider-qualified ID, bare canonical ID, or provider default:
fast-agent go --model tencent-eval.hy4-flash-dev --message "Review this code."
fast-agent go --model hy4-flash-dev --message "Review this code."
fast-agent go --model tencent-eval --message "Review this code."
Reasoning
hy4-flash-dev uses binary reasoning (high vs no_think), streams reasoning separately in reasoning_content, and defaults to high. Live sweeps confirmed that no_think and none disable reasoning while low, medium, high, xhigh, and max all enable reasoning without demonstrated ordered budgets, so fast-agent exposes binary high vs none/off (sent on the wire as reasoning_effort: "no_think"), with any other effort value falling back to high:
fast-agent go --model "tencent-eval.hy4-flash-dev?reasoning=high" --message "Solve this step by step."
fast-agent go --model "tencent-eval.hy4-flash-dev?reasoning=none" --message "Reply with only the final answer."
fast-agent go --model "tencent-eval.hy4-flash-dev?reasoning=off" --message "Reply with only the final answer."
fast-agent preserves raw reasoning_content across multi-turn tool-call replay for hy4-flash-dev.
Capabilities and behavior
- Streaming & usage: Supports SSE streaming (
stream: true) withstream_options: {"include_usage": true}and reports nestedcompletion_tokens_details.reasoning_tokens(already included incompletion_tokens). - Tool calling: Supports single and parallel function calls in both streaming and non-streaming modes. Note that the evaluation endpoint may still return multiple tool calls even when
parallel_tool_calls: falseis sent. - Structured output: Supports both
response_format: {"type": "json_schema", "json_schema": {..., "strict": true}}andresponse_format: {"type": "json_object"}. - Vision: Verified with inline base64 PNG data URLs (
image/png). - Prompt caching: Automatic server-side prefix caching (
usage.prompt_tokens_details.cached_tokens) is verified in both non-streaming and multi-turn streaming sessions (62.8%–94.5% cache hits observed in TB2.1 runs) and is normalized byfast-agentwhen reported. - Model discovery: The evaluation endpoint does not expose
GET /models; use the static model IDhy4-flash-dev.