GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Dratinia
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Dratinia
Model Tree
Qwen/Qwen3.5-2B
Input Modalities
Output Modalities
Supported Functionality
No evaluation data and no temporal hold-out problems were used for training.
All models were evaluated with one identical harness: non-thinking mode, the suffix above, and the pre-registered lite settings. TMH is a private temporal hold-out of 321 competition math problems published after the base model's release (JEE Main 2026 Session 2, JEE Advanced 2026, AIME 2026, HMMT February 2026). It is kept private so it stays out of training corpora; the repository publishes only its provenance manifest. Accuracy in %:
| model | TMH | TMH truncated | MATH-500 | GSM8K | GSM-Symbolic | D_dev |
|---|---|---|---|---|---|---|
| B0 (base Qwen3.5-2B) | 31.1 | 58.4 | 69.0 | 79.4 | 71.6 | - |
| A1 s1 | 30.5 | 47.4 | 66.2 | 81.1 | 74.0 | 71.0 |
| A1 s2 | 33.8 | 49.7 | 70.4 | 82.7 | 74.8 | 71.4 |
| A1 s3 (this adapter) | 29.4 | 51.4 | 70.8 | 82.3 | 76.2 | 70.2 |
| A1-S s1 | 32.9 | 51.6 | 69.4 | 83.2 | 75.2 | 72.4 |
Retention (lm-eval, lite settings, accuracy %):
| model | MMLU-Pro | IFEval | ARC-C | BBH | HumanEval+ | MBPP+ | mean |
|---|---|---|---|---|---|---|---|
| B0 (base Qwen3.5-2B) | 50.0 | 67.1 | 72.6 | 63.7 | 45.1 | 59.5 | 59.7 |
| A1 s1 | 50.1 | 67.7 | 74.3 | 65.3 | 47.6 | 57.7 | 60.5 |
| A1 s2 | 50.4 | 66.9 | 76.5 | 65.8 | 47.6 | 61.1 | 61.4 |
| A1 s3 (this adapter) | 50.1 | 66.5 | 75.4 | 65.2 | 49.4 | 59.0 | 61.0 |
| A1-S s1 | 49.0 | 64.1 | 76.1 | 65.9 | 45.7 | 59.3 | 60.0 |
Over A1's 3 seeds, the mean change from the base is +0.1 points on TMH, +0.1 on MATH-500, +2.6 on GSM8K, +3.4 on GSM-Symbolic and -8.9 points of TMH truncation. The six-task retention mean moves by under 2 points.
The study evaluated the adapters with vLLM LoRA serving (scripts/phase1_select.py in the repository):
python
from huggingface_hub import snapshot_downloadfrom vllm import LLM, SamplingParamsfrom vllm.lora.request import LoRARequest adapter = snapshot_download("Dratinia/qwen-reasoning-distill-a1-s3")llm = LLM("Qwen/Qwen3.5-2B", enable_lora=True, max_lora_rank=32, max_model_len=8192)msgs = [{"role": "user", "content": "What is 17 * 24?\n\nPlease reason step by step, and put your final answer within \\boxed{}."}]out = llm.chat(msgs, SamplingParams(temperature=0.6, top_p=0.95, top_k=20, max_tokens=4096), chat_template_kwargs={"enable_thinking": False}, lora_request=LoRARequest("qwen-reasoning-distill-a1-s3", 1, adapter))print(out[0].outputs[0].text)The adapter was trained with Unsloth and PEFT 0.21 on the Qwen3.5 module names (model.language_model.layers.*), so
PEFT loading in transformers needs the base model class used in training (Qwen3_5ForConditionalGeneration).
EVAL_PROTOCOL.md, git tag prereg-v1), amendments and all aggregate results:
https://github.com/Abhigyan1805/qwen-reasoning-distilladapter_model.safetensors sha256: d5ce5e817c10e55c611f5fe4b821dec5870faa22df5d612556e69d7c7a3442fafrom huggingface_hub import snapshot_downloadfrom vllm import LLM, SamplingParamsfrom vllm.lora.request import LoRARequest adapter = snapshot_download("Dratinia/qwen-reasoning-distill-a1-s3")llm = LLM("Qwen/Qwen3.5-2B", enable_lora=True, max_lora_rank=32, max_model_len=8192)msgs = [{"role": "user", "content": "What is 17 * 24?\n\nPlease reason step by step, and put your final answer within \\boxed{}."}]out = llm.chat(msgs, SamplingParams(temperature=0.6, top_p=0.95, top_k=20, max_tokens=4096), chat_template_kwargs={"enable_thinking": False}, lora_request=LoRARequest("qwen-reasoning-distill-a1-s3", 1, adapter))print(out[0].outputs[0].text)