Qwen3.8-27B-RFT-v1

Qwen3.8-27B after one round of rejection-sampling fine-tuning (RFT). The goal was shorter reasoning for the same accuracy.

Result in one line: accuracy is statistically the same as the base model in both thinking modes, and correct answers need about 10โ€“30% fewer reasoning tokens on most benchmarks.

A model by Horizon Tech d.o.o., Sarajevo, Bosnia and Herzegovina.

Model details

Developed by Horizon Tech d.o.o., Sarajevo, Bosnia and Herzegovina
Model type 27B-parameter vision-language model (text and images in, text out), with optional reasoning ("thinking") before the answer
Fine-tuned from Qwen/Qwen3.8-27B, revision 1d4bf0f
Method Rejection-sampling fine-tuning (RFT) with a LoRA adapter, merged into the weights
Weights bfloat16 safetensors, 18 shards, 52 GB; retrained MTP head included (mtp.* tensors)
Languages As the base model; evaluated in English and German
License Apache 2.0
Released November 2026
Author Hamza Abdagic (ORCID 0009-0002-4078-4922)
Contact hamza.abdagic@horizon-tech.io

Versions

Repository Format Size For
Qwen3.8-27B-RFT-v1 bfloat16 safetensors 52 GB Transformers, vLLM, SGLang; further fine-tuning
Qwen3.8-27B-RFT-v1-NVFP4 NVFP4 (4-bit) 20 GB vLLM on NVIDIA GPUs with FP4 support
Qwen3.8-27B-RFT-v1-mlx-4bit MLX 4-bit 14 GB Apple silicon, text only

What was done

  • Base: Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
  • Method: the base model answered training problems itself, programs graded the answers, and the model was trained on its own correct, shortest answers. No teacher model and no model-written grading.
  • Training: LoRA (rank 64, alpha 128), 1.5 epochs, learning rate 1e-4, sequences packed to 32,768 tokens. Validation loss went from 0.283 to 0.234.
  • These weights: the adapter merged into the base in fp32 and stored as bfloat16. The vision tower is unchanged.

Evaluation

Paired comparison against the base on the same items (McNemar test). No difference is significant (smallest p = 0.12 with thinking on).

Benchmark n Base, thinking on RFT v1, thinking on Base, thinking off RFT v1, thinking off
GSM8K 1319 97.95 97.88 96.59 96.82
MMLU-Pro (subset) 200 86.50 84.00 65.50 65.50
MATH-500 500 98.20 98.40 97.40 98.20
AIME 2025 30 83.33 83.33 70.00 63.33
HumanEval 164 93.90 95.73 97.56 97.56
MBPP 500 94.20 93.80 81.00 80.80
IFEval 541 89.09 90.76 82.07 81.89
German MMLU (subset) 200 83.50 86.50 80.00 81.00

Reasoning tokens per correct answer, thinking on (RFT v1 vs base, same run as the table):

Benchmark Base RFT v1 Change
GSM8K 389 338 โˆ’13%
MMLU-Pro 1,702 1,179 โˆ’31%
MATH-500 1,481 1,328 โˆ’10%
AIME 2025 9,018 8,845 โˆ’2%
HumanEval 1,531 1,070 โˆ’30%
MBPP 2,885 2,603 โˆ’10%

The number of answers cut off at the token limit did not change in a consistent direction.

Limitations

  • This is not a more accurate model than the base. Several of these benchmarks are close to their ceiling for the base, so they cannot show a gain.
  • The multi-token-prediction (MTP) head was retrained for this model by self-distillation (the main weights are unchanged by it). Measured on 120 held-out documents, the retrained head gets 3.15 drafted tokens accepted at 3 drafts and 4.06 at 5 drafts; the head copied from the base got 3.02 and 3.63. These are offline numbers, not serving speed.
  • Scores come from our own harness and are meant for the base-vs-RFT comparison, not for comparison with numbers published elsewhere.

How to use

Same usage as the base model (Transformers, vLLM, SGLang); the chat template and thinking controls are unchanged.

Transformers

import torch
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

repo = "ghostrider0470/Qwen3.8-27B-RFT-v1"
tok = AutoTokenizer.from_pretrained(repo)
model = Qwen3_5ForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "A train leaves at 14:10 and arrives at 16:45. How long is the trip?"}]
prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=True)  # False = answer without reasoning
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=2048)
print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))

vLLM

vllm serve ghostrider0470/Qwen3.8-27B-RFT-v1

Intended use

  • A drop-in replacement for Qwen3.8-27B where reasoning length matters: the same answers for fewer reasoning tokens, so lower latency and cost per answer with thinking on.
  • A starting point for further fine-tuning.
  • It has the same safety behaviour and the same limits as the base model. It was not trained on any user or customer data.

Citation

@misc{horizontech2026rftv1,
  author       = {Abdagic, Hamza},
  title        = {Qwen3.8-27B-RFT-v1},
  year         = {2026},
  publisher    = {Hugging Face},
  organization = {Horizon Tech d.o.o., Sarajevo},
  howpublished = {\url{https://huggingface.co/ghostrider0470/Qwen3.8-27B-RFT-v1}}
}

Acknowledgements

License

Apache 2.0, as the base model.

Downloads last month
95
Safetensors
Model size
28B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ghostrider0470/Qwen3.8-27B-RFT-v1

Base model

Qwen/Qwen3.8-27B
Finetuned
(957)
this model
Quantizations
2 models