SRIQwen-V1.3
SRIQwen-V1.3 is a language model by SRIQ, supervised fine-tuned from Qwen/Qwen3.8-27B. Our focus is shorter reasoning traces while preserving answer quality.
The name combines SRIQ and Qwen, sharing the Q.
| Base model | Qwen/Qwen3.8-27B |
| Method | LoRA supervised fine-tuning, merged into the released weights |
| Dataset | sriq-ai/sriq-sft-v1.3 |
| Training context | 131,072 tokens |
| Precision | bfloat16 |
Training
Supervised fine-tuning on sriq-ai/sriq-sft-v1.3, a ShareGPT-format set whose assistant targets pair compressed Simplified Chinese reasoning with the original final answer. The compression step never rewrites the final answer, so the model learns to shorten the trace rather than the result.
Training used Unsloth with a LoRA adapter over the 4-bit base checkpoint.
Hyperparameters
| Setting | Value |
|---|---|
LoRA rank (r) | 64 |
| LoRA alpha | 128 |
| Learning rate | 2e-4 |
| Steps | 103 |
| Max sequence length | 131,072 |
| Final training loss | ≈ 0.3 |
Target modules are the Unsloth/PEFT defaults — attention and MLP projections, no embedding or head tuning:
q_proj k_proj v_proj o_proj gate_proj up_proj down_proj
The adapter is merged into the weights published here, so no PEFT adapter is needed at inference time — load the repo directly.
Evaluation
No benchmarks were run for this release. SRIQwen-V1.3 ships unmeasured: the claims above describe the training setup, not verified gains in accuracy, token use, or response time. Evaluate it on your own workload before relying on it.
Usage
The base model supports a 262,144-token context. Fine-tuning used 131,072-token sequences, so behaviour beyond that length is untested.
vLLM
vllm serve sriq-ai/Sriqwen-V1.3 \
--served-model-name Sriqwen-V1.3 \
--host 0.0.0.0 --port 8000 \
--tensor-parallel-size 1 \
--dtype bfloat16 \
--kv-cache-dtype auto \
--max-model-len 262144 \
--max-num-seqs 4 \
--max-num-batched-tokens 4096 \
--gpu-memory-utilization 0.90 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder
This serves the vision tower alongside the language model, so the endpoint accepts image
inputs. Add --language-model-only to disable vision and free the memory it occupies.
Transformers
The checkpoint is a Qwen3_5ForConditionalGeneration, so load it with AutoProcessor and
AutoModelForImageTextToText — not AutoModelForCausalLM, which resolves to the text-only
class for this architecture.
from transformers import AutoProcessor, AutoModelForImageTextToText
model_id = "sriq-ai/Sriqwen-V1.3"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id, dtype="bfloat16", device_map="auto"
)
messages = [{"role": "user", "content": [{"type": "text", "text": "Why is the sky blue?"}]}]
inputs = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Reasoning is emitted inside <think> tags in Simplified Chinese, followed by the final answer
in the language of the prompt.
Credits
Developed by SRIQ. Base model by the Qwen team.