README.md

metadata

license: mit
language:
  - en
base_model:
  - WeiboAI/VibeThinker-3B
base_model_relation: quantized
tags:
  - math
  - code
  - reasoning
  - quantized
  - gptq
  - w4a16
  - compressed-tensors
  - vllm
pipeline_tag: text-generation
library_name: transformers

VibeThinker-3B-W4A16 (4-bit GPTQ, group size 128)

A 4-bit weight-quantized version of WeiboAI/VibeThinker-3B, produced so the model fits and runs on a 6 GB consumer GPU.

TL;DR — The original BF16 release is 5.8 GB; its weights alone don't fit in 6 GB of VRAM. Quantized to W4A16 (INT4, group size 128) with llmcompressor, the weights shrink to **2.0 GB** and the model serves happily in vLLM on an RTX 3050 6 GB Laptop GPU at ~67 tok/s — with the step-by-step math reasoning intact.

The story

I wanted to run VibeThinker-3B, a small but strong math/STEM reasoning model, on a laptop with an RTX 3050 (6 GB VRAM, 50 W). The problem:

Format	Weights	Fits 6 GB?
BF16 (original)	~5.8 GB	❌ weights alone fill the card; nothing left for KV cache + overhead
W4A16 (this repo)	~2.0 GB	✅ leaves ~2.5 GB for KV cache (≈74k tokens)

So I quantized it locally, then loaded the result back into vLLM and confirmed it produces correct, fully-reasoned answers — all on the 6 GB laptop GPU.

What was done

• Method: GPTQ, W4A16 scheme (4-bit symmetric INT weights, FP16 activations), group size 128, lm_head left unquantized.
• Tool: llmcompressor 0.12, output in compressed-tensors (pack-quantized) format.
• Calibration: 512 samples from HuggingFaceH4/ultrachat_200k @ 2048 tokens.
• Hardware used for quantization: RTX 3050 6 GB — the model was loaded on CPU and quantized layer-by-layer (sequential GPU on-loading), keeping peak VRAM ~3.2 GB.

Measured results (RTX 3050 6 GB, vLLM 0.22)

Metric	Value
Weights on GPU	1.99 GiB
KV cache available	2.55 GiB (74,144 tokens)
Max concurrency @ 8192 ctx	9.05×
Throughput	~67 tok/s (eager)
Kernel	Marlin W4A16 + FlashAttention 2

Usage (vLLM)

vllm serve syedazeez/VibeThinker-3B-W4A16-G128 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.92 \
  --dtype float16

from vllm import LLM, SamplingParams

llm = LLM(model="syedazeez/VibeThinker-3B-W4A16-G128",
          max_model_len=8192, gpu_memory_utilization=0.92, dtype="float16")
out = llm.chat(
    [[{"role": "user", "content": "What is the remainder when 7^100 is divided by 13?"}]],
    SamplingParams(temperature=0.6, max_tokens=1500),
)
print(out[0].outputs[0].text)

This is a reasoning model: it emits a <think>...</think> trace before the final answer, so give it a generous max_tokens (≥1500) or the answer may be truncated.

Notes & limitations

• Quantization is lossy; on hard benchmarks expect a small quality drop vs the BF16 original. For most math/STEM prompts the reasoning quality is preserved in informal testing.
• Inherits the base model's scope: tuned for verifiable tasks (competition math, coding, STEM). Not trained for tool-calling/agents or broad open-domain knowledge.
• KV cache fits ~74k tokens at this config, so --max-model-len can be pushed well above 8192 on 6 GB.

License & attribution

This derivative is released under the MIT License, the same license as the original WeiboAI/VibeThinker-3B. All credit for the model itself goes to the original authors. Only the 4-bit quantization was performed here.

@misc{xu2026vibethinker3bexploringfrontierverifiable,
      title={VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models},
      author={Sen Xu and Shixi Liu and Wei Wang and Jixin Min and Yingwei Dai and Zhibin Yin and Yirong Chen and Xin Zhou and Junlin Zhang},
      year={2026},
      eprint={2606.16140},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2606.16140}
}