GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ConwayResearch
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ConwayResearch
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
All numbers below come from one harness, run on Qwen3.5-4B and on this model with identical settings: thinking on, temperature 0.6, top_p 0.95, top_k 20, no presence penalty, and at most 32,768 output tokens. An answer that runs past the budget counts as wrong. We report these settings instead of Qwen's card settings (temperature 1.0 with a presence penalty of 1.5), so the numbers here are not comparable with Qwen's published figures.
| Benchmark | Qwen3.5-4B | Underdog Poodle 4B |
|---|---|---|
| BFCL-V4, official harness | 56.8 | 65.3 |
| BFCL-120 (frozen 120-case subset of BFCL, 3 seeds) | 77.5 | 80.6 |
| IFEval (prompt-level, loose) | 60.6 | 85.4 |
| IFBench (prompt-level, loose) | 33.3 | 38.7 |
| Parallel calls (100 cases) | 69 | 77 |
| WebWalkerQA (150 questions, LLM judge) | 30.4 | 31.1 |
| τ³-bench average (4 domains) | 57.9 | 60.7 |
BFCL-V4 by category:
| Category | Qwen3.5-4B | Underdog Poodle 4B |
|---|---|---|
| Non-live AST | 85.6 | 86.1 |
| Live | 81.0 | 79.1 |
| Multi-turn | 77.8 | 82.3 |
| Web search | 14.0 | 30.5 |
| Memory | 29.3 | 48.6 |
| Relevance detection | 87.5 | 81.3 |
| Irrelevance detection | 81.4 | 82.6 |
τ³-bench runs 2 trials per task with a 131,072-token context. The user simulator and the assertion judge are GPT-5.4 mini, the τ³ package defaults. An episode that fails to produce an official reward counts as a failure for both models.
BFCL-V4 uses the official prompts, checkers and score sheet. The one change is the web-search backend: searches go through OpenRouter's web plugin instead of SerpAPI, and every result is cached so both models saw the same pages.
The model also finishes far more of its answers inside the token budget. On IFEval, 98% of its answers finish against 63% for Qwen3.5-4B; on IFBench, 79% against 43%.
Tested with vLLM 0.32.0 (Python 3.12, NVIDIA H100). The model uses Qwen3.5's chat template and tool-call format:
bash
pip install vllm==0.32.0vllm serve ConwayResearch/Underdog-Poodle-4B \ --tool-call-parser qwen3_coder --reasoning-parser qwen3 --enable-auto-tool-choiceKeep thinking on (the default) and use temperature 0.6, top_p 0.95, top_k 20:
bash
curl -s http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{ "model": "ConwayResearch/Underdog-Poodle-4B", "messages": [{"role": "user", "content": "Where is my order 48213?"}], "tools": [{"type": "function", "function": { "name": "get_order_status", "description": "Look up the shipping status of an order.", "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}}, "required": ["order_id"]} }}], "temperature": 0.6, "top_p": 0.95, "top_k": 20}'Apache-2.0, inherited from Qwen3.5-4B. Post training by ConwayResearch; see NOTICE.
pip install vllm==0.32.0vllm serve ConwayResearch/Underdog-Poodle-4B \ --tool-call-parser qwen3_coder --reasoning-parser qwen3 --enable-auto-tool-choicecurl -s http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{ "model": "ConwayResearch/Underdog-Poodle-4B", "messages": [{"role": "user", "content": "Where is my order 48213?"}], "tools": [{"type": "function", "function": { "name": "get_order_status", "description": "Look up the shipping status of an order.", "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}}, "required": ["order_id"]} }}], "temperature": 0.6, "top_p": 0.95, "top_k": 20}'