[Model Submission] OpenAI-Compatible API — Xiaothink-T17-RWKV5-MLA-0.6B-Frontend

[Ericsjq]

Organization: Xiaothink Framework
Submitter: Individual developer / open-source contributor
Model type: Open weights + public API (OpenAI-compatible)

We would like to submit Xiaothink-T17-RWKV5-MLA-0.6B-Frontend to the Arena for evaluation. It is a 0.6B-parameter model trained from scratch, specialized in generating beautiful static frontend pages (single-file HTML/CSS/JS) from ultra-short prompts (10–20 tokens).

---

Why This Model Fits Arena

• Vertical Specialization: Trained specifically for frontend static-page generation — a narrow but high-traffic category on Code Arena WebDev.
• Primary Language Support (Chinese): The model is specifically optimized for Chinese prompts and generating frontend pages with Chinese content, filling a gap for Chinese-language frontend tasks in the Arena.
• Ultra-short Prompt Competence: Produces complete, production-quality single-file HTML from ≤20-token descriptions, aligning with real user prompt patterns (Landing Page, Portfolio, Dashboard, etc.).
• Competitive Blind-test Performance: In an independent blind evaluation judged by DeepSeek-V4, the model outperforms Llama-3-8B and approaches GPT-OSS-20B quality on static frontend page generation tasks.
• Open Weights: Full model weights publicly available on ModelScope for transparency and reproducibility.
• Novel Architecture: RWKV-5 with Multi-head Latent Attention (MLA) — an RNN-Transformer hybrid offering constant memory usage and stable inference cost, relevant to Arena's diverse model pool.

---

API Specification

Field	Value
Base URL	http://cn-hk-bgp-4.ofalias.net:27076/llmapi/v1
Protocol	OpenAI API v1 compatible
Endpoints	GET /models, POST /chat/completions
Auth	Authorization: Bearer <api_key>
Model name (for requests)	t17-rwkv5-mla-0.6B-frontend
Streaming	Supported (stream: true → SSE; stream: false → JSON)
Output format	WanlyFrontend Code + Single-file HTML code

---

Example Requests

Non-streaming:

curl -X POST http://cn-hk-bgp-4.ofalias.net:27076/llmapi/v1/chat/completions \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "t17-rwkv5-mla-0.6B-frontend",
    "messages": [{"role": "user", "content": "极简的个人作品集网站"}],
    "stream": false
  }'

Streaming (SSE):

curl -N -X POST http://cn-hk-bgp-4.ofalias.net:27076/llmapi/v1/chat/completions \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "t17-rwkv5-mla-0.6B-frontend",
    "messages": [{"role": "user", "content": "加密货币数据仪表盘"}],
    "stream": true
  }'

Model list:

curl -X GET http://cn-hk-bgp-4.ofalias.net:27076/llmapi/v1/models \
  -H "Authorization: Bearer <your-api-key>"

---

Obtaining an Evaluation API Key

The API is not open-access (rate-limited to protect inference compute). The Arena team can obtain a dedicated evaluation key with elevated quota by contacting:

• Email: xiaothink@foxmail.com
• Subject: [Arena Evaluation Key Request] Xiaothink-T17 Frontend
• We will provision a long-lived key with sufficient rate limits for Arena's voting volume and commit to keeping it active for the full 30-day evaluation window per Arena's Leaderboard Policy.

---

Open Weights & Technical Details

• ModelScope: https://www.modelscope.cn/models/ericsjq/Xiaothink-T17-RWKV5-MLA-0.6B-Frontend
• Official site: http://xiaothink.top/
• Architecture: RWKV-5 with Multi-head Latent Attention (MLA)
• Parameters: 0.6B
• Primary Language: Chinese (optimized for Chinese prompts and content generation)
• License: see ModelScope model card
• Training: from-scratch (not a fine-tune of an existing base model)
• Specialty: static frontend page generation (single-file HTML/CSS/JS) from ultra-short prompts

---

Blind-Test Evidence (DeepSeek-V4 as Judge)

An independent pairwise blind evaluation was conducted using DeepSeek-V4 as the LLM judge, comparing generated static frontend pages on a held-out prompt set. Results:

Model	Relative Quality
Llama-3-8B (general)	Baseline — Xiaothink-T17 outperforms
GPT-OSS-20B (general)	Xiaothink-T17 approaches parity
Xiaothink-T17-0.6B-Frontend	Exceeds Llama3-8B, ~matches GPT-OSS-20B in this narrow vertical

Full evaluation logs and prompt sets are available on request. We understand Arena will re-evaluate using its own native mechanism (human blind voting); the DeepSeek-V4 score is provided only as a capability reference.

---

Availability Commitment

• API endpoint will remain publicly accessible for at least 30 days post-launch per Arena Leaderboard Policy.
• We will not pause or revoke access once testing begins.
• Compute is self-hosted on dedicated GPU infrastructure — not a shared free tier.
• We are prepared for the ~200-vote stabilization period and will maintain capacity throughout.

---

Preferred Arena Category

• Primary: Code Arena → WebDev → HTML / Website
• Secondary: Code Arena → WebDev → UI Component

---

Contact

• Email: xiaothink@foxmail.com
• Organization: Xiaothink Framework
• Website: http://xiaothink.top/

We are happy to make any API adjustments (additional endpoints, response format tweaks, etc.) the Arena team requires. Looking forward to your feedback.