--- license: mit language: - en base_model: - Qwen/Qwen2.5-Coder-3B tags: - math - code - reasoning - gpqa - instruction-following - heretic - uncensored - decensored - abliterated pipeline_tag: text-generation library_name: transformers --- # This is a decensored version of a model, made using [Heretic](https://heretic-project.org) v1.4.0 ## Abliteration parameters | Parameter | Value | | :-------- | :---: | | **direction_index** | 23.30 | | **attn.o_proj.max_weight** | 1.26 | | **attn.o_proj.max_weight_position** | 21.83 | | **attn.o_proj.min_weight** | 1.20 | | **attn.o_proj.min_weight_distance** | 13.99 | | **mlp.down_proj.max_weight** | 1.49 | | **mlp.down_proj.max_weight_position** | 21.32 | | **mlp.down_proj.min_weight** | 0.86 | | **mlp.down_proj.min_weight_distance** | 13.56 | ## Performance | Metric | This model | Original model (a model) | | :----- | :--------: | :---------------------------: | | **KL divergence** | 0.0255 | 0 *(by definition)* | | **Refusals** | 9/100 | 64/100 | ## Residual Geometry | Layer | S(g,b) | S(g*,b*) | S(g,r) | S(g*,r*) | S(b,r) | S(b*,r*) | \|g\| | \|g*\| | \|b\| | \|b*\| | \|r\| | \|r*\| | Silh | | ----- | ------ | -------- | ------- | -------- | ------ | -------- | ------ | ------ | ------ | ------ | ----- | ------ | ------ | | 1 | 0.9994 | 0.9994 | 0.1419 | 0.1613 | 0.1773 | 0.1966 | 12.07 | 12.06 | 12.14 | 12.14 | 0.44 | 0.44 | 0.0582 | | 2 | 0.9994 | 0.9994 | 0.1141 | 0.1236 | 0.1484 | 0.1577 | 15.81 | 15.80 | 15.88 | 15.88 | 0.55 | 0.55 | 0.0670 | | 3 | 0.9987 | 0.9987 | 0.0273 | 0.0306 | 0.0785 | 0.0816 | 16.28 | 16.28 | 16.33 | 16.33 | 0.84 | 0.83 | 0.0788 | | 4 | 0.9979 | 0.9979 | 0.0460 | 0.0506 | 0.1110 | 0.1152 | 16.28 | 16.28 | 16.36 | 16.36 | 1.07 | 1.06 | 0.0868 | | 5 | 0.9973 | 0.9974 | 0.1027 | 0.1029 | 0.1750 | 0.1743 | 16.83 | 16.83 | 17.00 | 17.01 | 1.25 | 1.23 | 0.0941 | | 6 | 0.9975 | 0.9976 | 0.0628 | 0.0623 | 0.1332 | 0.1319 | 17.63 | 17.63 | 17.75 | 17.76 | 1.26 | 1.24 | 0.0956 | | 7 | 0.9964 | 0.9964 | -0.0058 | -0.0099 | 0.0794 | 0.0748 | 18.16 | 18.18 | 18.22 | 18.23 | 1.55 | 1.54 | 0.0870 | | 8 | 0.9941 | 0.9941 | 0.0083 | 0.0045 | 0.1172 | 0.1131 | 18.91 | 18.93 | 19.04 | 19.06 | 2.07 | 2.07 | 0.0952 | | 9 | 0.9849 | 0.9852 | -0.0096 | -0.0132 | 0.1636 | 0.1582 | 19.40 | 19.44 | 19.66 | 19.68 | 3.40 | 3.37 | 0.1217 | | 10 | 0.9854 | 0.9857 | -0.0042 | -0.0089 | 0.1660 | 0.1596 | 20.51 | 20.56 | 20.80 | 20.83 | 3.54 | 3.51 | 0.1232 | | 11 | 0.9869 | 0.9871 | 0.0023 | -0.0012 | 0.1637 | 0.1587 | 20.92 | 20.97 | 21.20 | 21.23 | 3.42 | 3.39 | 0.1262 | | 12 | 0.9860 | 0.9862 | -0.0215 | -0.0283 | 0.1455 | 0.1377 | 25.71 | 25.78 | 25.98 | 26.02 | 4.33 | 4.31 | 0.1191 | | 13 | 0.9838 | 0.9839 | -0.1014 | -0.1188 | 0.0786 | 0.0605 | 25.86 | 25.96 | 25.80 | 25.82 | 4.65 | 4.65 | 0.1191 | | 14 | 0.9837 | 0.9839 | -0.0898 | -0.1108 | 0.0906 | 0.0685 | 26.00 | 26.12 | 26.00 | 26.02 | 4.69 | 4.68 | 0.1131 | | 15 | 0.9830 | 0.9833 | -0.0709 | -0.0856 | 0.1133 | 0.0971 | 26.07 | 26.18 | 26.17 | 26.21 | 4.81 | 4.79 | 0.1106 | | 16 | 0.9808 | 0.9813 | 0.0049 | -0.0102 | 0.1995 | 0.1825 | 26.02 | 26.17 | 26.56 | 26.61 | 5.17 | 5.13 | 0.1142 | | 17 | 0.9858 | 0.9862 | 0.0985 | 0.0796 | 0.2640 | 0.2438 | 29.96 | 30.13 | 30.91 | 30.97 | 5.21 | 5.15 | 0.1160 | | 18 | 0.9857 | 0.9861 | 0.0173 | 0.0000 | 0.1854 | 0.1664 | 29.06 | 29.21 | 29.57 | 29.63 | 4.98 | 4.93 | 0.1078 | | 19 | 0.9863 | 0.9867 | 0.0289 | 0.0055 | 0.1935 | 0.1679 | 31.16 | 31.35 | 31.75 | 31.80 | 5.24 | 5.17 | 0.1101 | | 20 | 0.9856 | 0.9860 | 0.0036 | -0.0164 | 0.1727 | 0.1508 | 30.69 | 30.85 | 31.16 | 31.20 | 5.27 | 5.21 | 0.1071 | | 21 | 0.9810 | 0.9813 | 0.1356 | 0.1211 | 0.3253 | 0.3100 | 36.41 | 36.57 | 38.15 | 38.19 | 7.47 | 7.41 | 0.1585 | | 22 | 0.9789 | 0.9792 | 0.1084 | 0.0922 | 0.3092 | 0.2923 | 38.37 | 38.60 | 40.11 | 40.19 | 8.24 | 8.19 | 0.1662 | | 23 | 0.9724 | 0.9727 | 0.0068 | -0.0035 | 0.2401 | 0.2288 | 37.52 | 37.71 | 38.65 | 38.74 | 9.02 | 8.99 | 0.1715 | | 24 | 0.9685 | 0.9685 | -0.0140 | -0.0286 | 0.2355 | 0.2212 | 40.01 | 40.24 | 41.17 | 41.24 | 10.25 | 10.27 | 0.1857 | | 25 | 0.9549 | 0.9544 | -0.1280 | -0.1385 | 0.1722 | 0.1636 | 42.07 | 42.28 | 42.35 | 42.44 | 12.68 | 12.80 | 0.2114 | | 26 | 0.9512 | 0.9504 | -0.1326 | -0.1426 | 0.1798 | 0.1723 | 44.42 | 44.64 | 44.76 | 44.85 | 13.94 | 14.09 | 0.2190 | | 27 | 0.9524 | 0.9513 | -0.1822 | -0.1915 | 0.1263 | 0.1202 | 51.97 | 52.19 | 51.51 | 51.60 | 15.97 | 16.20 | 0.2333 | | 28 | 0.9331 | 0.9309 | -0.1472 | -0.1589 | 0.2182 | 0.2126 | 53.06 | 53.29 | 53.77 | 53.85 | 19.55 | 19.92 | 0.2326 | | 29 | 0.9280 | 0.9251 | -0.1168 | -0.1313 | 0.2617 | 0.2551 | 61.41 | 61.77 | 63.19 | 63.33 | 23.71 | 24.26 | 0.2238 | | 30 | 0.9122 | 0.9090 | -0.1673 | -0.1811 | 0.2514 | 0.2452 | 66.16 | 66.62 | 67.40 | 67.58 | 28.01 | 28.64 | 0.2210 | | 31 | 0.9066 | 0.9034 | -0.1648 | -0.1825 | 0.2669 | 0.2567 | 82.24 | 83.01 | 84.17 | 84.45 | 36.02 | 36.82 | 0.2197 | | 32 | 0.9124 | 0.9102 | -0.0878 | -0.1084 | 0.3275 | 0.3132 | 112.01 | 113.23 | 118.09 | 118.53 | 48.51 | 49.39 | 0.2145 | | 33 | 0.9271 | 0.9252 | -0.1205 | -0.1459 | 0.2604 | 0.2405 | 146.74 | 148.56 | 150.88 | 151.41 | 56.96 | 58.08 | 0.2068 | | 34 | 0.9447 | 0.9435 | -0.0870 | -0.1083 | 0.2446 | 0.2272 | 177.54 | 179.19 | 182.41 | 182.92 | 60.06 | 60.97 | 0.2024 | | 35 | 0.9363 | 0.9350 | -0.1146 | -0.1319 | 0.2415 | 0.2282 | 180.56 | 182.04 | 184.84 | 185.34 | 65.32 | 66.30 | 0.2059 | | 36 | 0.8726 | 0.8711 | -0.1135 | -0.1285 | 0.3862 | 0.3750 | 86.94 | 87.92 | 93.65 | 94.06 | 46.04 | 46.57 | 0.1981 | g = mean of residual vectors for good prompts g* = geometric median of residual vectors for good prompts b = mean of residual vectors for bad prompts b* = geometric median of residual vectors for bad prompts r = refusal direction for means (i.e., b - g) r* = refusal direction for geometric medians (i.e., b* - g*) S(x,y) = cosine similarity of x and y |x| = L2 norm of x Silh = Mean silhouette coefficient of residuals for good/bad clusters ## Residual Vectors Visualization ![VibeThinker-3B-animation](pictures/animation-VibeThinker-3B.gif) ----- # VibeThinker-3B

GitHub  |  Hugging Face  |  Technical Report

## Introduction VibeThinker-3B is a further exploration of the VibeThinker series at the 3B-parameter scale, focusing on challenging reasoning tasks with clear verification signals, such as mathematics, coding, and STEM. By systematically optimizing the Spectrum-to-Signal Principle (SSP) post-training pipeline introduced in VibeThinker-1.5B, VibeThinker-3B achieves strong performance on AIME, HMMT, IMO-AnswerBench, LiveCodeBench, and recent LeetCode contests, reaching the performance range of top-tier frontier reasoning models, including Qwen3.6 Plus, Gemini 3 Pro, GLM-5, and Kimi K2.5, on verifiable reasoning benchmarks. Motivated by these observations, we propose the Parametric Compression-Coverage Hypothesis: different capabilities depend on model parameters in fundamentally different ways. Verifiable reasoning is closer to a highly compressible, parameter-dense capability, centered on multi-step reasoning, constraint satisfaction, self-correction, and answer verification. When the task space is sufficiently structured and feedback signals are sufficiently reliable, compact models may also carry near-frontier reasoning capabilities. In contrast, open-domain knowledge, general-purpose dialogue, and long-tail scenario understanding rely more heavily on large-scale parameters to broadly cover facts, concepts, and world knowledge. From VibeThinker-1.5B to VibeThinker-3B, our goal is not to build a small model that replaces large-scale models, but to examine the real boundaries of small models along specific capability dimensions. With VibeThinker-3B, we aim to show that small models should not be viewed merely as a compromise for reducing deployment costs. For capability domains with clear feedback and verification mechanisms, SLMs emerge as a promising research trajectory toward frontier-level performance that is fundamentally complementary to the traditional parameter scaling paradigm. ![alt text](pictures/Abstrct.png) ## Key Performance Data 📏 In terms of reasoning accuracy relative to model scale, VibeThinker-3B reaches 76.4 on IMO-AnswerBench, a highly challenging benchmark with 400 IMO-level problems, with only 3B parameters, and improves to 80.6 with Claim-Level Reliability Assessment (CLR), a test-time scaling strategy for answer-verifiable reasoning tasks. This demonstrates that a model within a strictly small-model regime can reach the performance range of substantially larger models, such as DeepSeek V3.2 (78.3, 671B), GLM-5 (82.5, 744B), and Kimi K2.5 (81.8, 1T). ![alt text](pictures/Acc_and_Scale.png) 💡 VibeThinker-3B achieves strong results across mathematics, coding, knowledge, and instruction-following benchmarks. ![alt text](pictures/VibeThiinker-3B.png) 🔁 VibeThinker-3B achieves competitive results against first-tier reasoning models and reaches the performance range of top-tier systems on several verifiable reasoning benchmarks. ![alt text](pictures/VibeThinker-3B+CLR.png) 🏆 To further test the model's out-of-distribution performance, we evaluate VibeThinker-3B on recent unseen LeetCode weekly and biweekly contests (Python) from Apr. 25 to May 31, 2026. VibeThinker-3B passes **123/128** first-attempt submissions, corresponding to a **96.1%** acceptance rate. ![alt text](pictures/LeetCode.png) ## Training Pipeline VibeThinker-3B follows the **Spectrum-to-Signal Principle (SSP)** introduced in VibeThinker-1.5B. The SFT stage constructs a broad spectrum of valid reasoning trajectories, while the RL stage amplifies correct reasoning signals using verifiable rewards. ![alt text](pictures/Architecture.png) The training pipeline contains the following stages: 1. **Curriculum-based two-stage SFT** - Stage 1 focuses on broad capability coverage across math, code, STEM reasoning, general dialogue, and instruction following. - Stage 2 shifts toward harder and longer-horizon reasoning samples. - Diversity-Exploring Distillation is used to preserve multiple valid solution paths. 2. **Multi-domain Reasoning RL** - VibeThinker-3B reuses MaxEnt-Guided Policy Optimization (MGPO). - RL is applied sequentially to math, code, and STEM reasoning tasks. - Training uses a single 64K long-context window to preserve complete long-horizon reasoning trajectories. 3. **Offline Self-Distillation** - High-quality trajectories from Math, Code, and STEM RL checkpoints are filtered and distilled back into a unified student model. - A learning-potential score is used to prioritize traces that are correct but not yet well modeled by the student. 4. **Instruct RL** - The final stage improves controllability on user-facing prompts. - Rule-based validators and rubric-based reward models are used for format-sensitive and open-ended instruction data. ## Usage Guidelines We recommend using VibeThinker-3B for competitive-style math, coding, STEM reasoning, and other tasks where the target answer can be verified. For broad open-domain knowledge tasks, larger general-purpose models may still be more suitable. For benchmark-style evaluation, the technical report uses vLLM with: - `temperature=1.0` - `top_p=0.95` - `top_k=-1` ## Quick Start Required: **transformers>=4.54.0** Recommended for better inference performance: **vLLM==0.10.1 or SGLang>=0.4.9.post6** ```python from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig class VibeThinker: def __init__(self, model_path): self.model_path = model_path self.model = AutoModelForCausalLM.from_pretrained( self.model_path, low_cpu_mem_usage=True, torch_dtype="bfloat16", device_map="auto", ) self.tokenizer = AutoTokenizer.from_pretrained( self.model_path, trust_remote_code=True, ) def infer_text(self, prompt): messages = [{"role": "user", "content": prompt}] text = self.tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, ) model_inputs = self.tokenizer([text], return_tensors="pt").to(self.model.device) generation_config = dict( max_new_tokens=102400, do_sample=True, temperature=1.0, top_p=0.95, top_k=None, ) generated_ids = self.model.generate( **model_inputs, generation_config=GenerationConfig(**generation_config), ) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) ] return self.tokenizer.batch_decode( generated_ids, skip_special_tokens=True, )[0] if __name__ == "__main__": model = VibeThinker("WeiboAI/VibeThinker-3B") prompt = "Your Prompt" print(model.infer_text(prompt)) ``` ## License The model repository is licensed under the MIT License. ## Citations & References If you use VibeThinker-3B in your research or product, please cite: ```bibtex @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}, } ```