--- license: mit language: - en - ru base_model: WeiboAI/VibeThinker-3B tags: - math - code - reasoning - gpqa - instruction-following - gguf - llama.cpp pipeline_tag: text-generation --- # KakTakOne/VibeThinker-3B-GGUF This repository contains GGUF format model files for [WeiboAI/VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B). VibeThinker-3B is a 3-billion-parameter dense reasoning model designed for verifiable reasoning tasks like mathematics, competitive programming, and STEM. ## Available Quantizations | File | Quantization | Size | Description | | --- | --- | --- | --- | | **[VibeThinker-3B-f16.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-f16.gguf)** | FP16 | 6.18 GB | Original precision, best quality. | | **[VibeThinker-3B-Q8_0.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-Q8_0.gguf)** | Q8_0 | 3.29 GB | High quality, recommended for resource-rich environments. | | **[VibeThinker-3B-Q5_K_M.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-Q5_K_M.gguf)** | Q5_K_M | 2.22 GB | Good balance between size and quality. | | **[VibeThinker-3B-Q4_K_M.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-Q4_K_M.gguf)** | Q4_K_M | 1.93 GB | Fastest, lightweight (most popular for daily use). | --- ## 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. ## Key Performance Data * πŸ“ 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. 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). * πŸ† To further test the model's out-of-distribution performance, it was evaluated 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. ## Training Pipeline VibeThinker-3B follows the **Spectrum-to-Signal Principle (SSP)**. The SFT stage constructs a broad spectrum of valid reasoning trajectories, while the RL stage amplifies correct reasoning signals using verifiable rewards: 1. **Curriculum-based two-stage SFT** (Stage 1: broad capability coverage, Stage 2: harder/longer samples). 2. **Multi-domain Reasoning RL** using MaxEnt-Guided Policy Optimization (MGPO) with a 64K context window. 3. **Offline Self-Distillation** using a learning-potential score to distill high-quality trajectories back into a student model. 4. **Instruct RL** to improve format controllability on user-facing prompts. --- ## How to use You can load these GGUF files in **LM Studio**, **Ollama**, **llama.cpp**, or any other GGUF-compatible inference engine. ### LM Studio Search for `KakTakOne/VibeThinker-3B-GGUF` directly in LM Studio search bar and download the desired quantization. ### CLI (llama.cpp) ```bash llama-cli -m VibeThinker-3B-Q4_K_M.gguf -p "2+2=" -n 128 ``` ## Citations & References ```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}, } ``` ---
Π§ΠΈΡ‚Π°Ρ‚ΡŒ описаниС Π½Π° русском языкС (Russian Description) # KakTakOne/VibeThinker-3B-GGUF Π’ этом Ρ€Π΅ΠΏΠΎΠ·ΠΈΡ‚ΠΎΡ€ΠΈΠΈ содСрТатся Ρ„Π°ΠΉΠ»Ρ‹ ΠΌΠΎΠ΄Π΅Π»Π΅ΠΉ Π² Ρ„ΠΎΡ€ΠΌΠ°Ρ‚Π΅ GGUF для [WeiboAI/VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B). VibeThinker-3B β€” это модСль рассуТдСний (reasoning model) с 3 ΠΌΠΈΠ»Π»ΠΈΠ°Ρ€Π΄Π°ΠΌΠΈ ΠΏΠ°Ρ€Π°ΠΌΠ΅Ρ‚Ρ€ΠΎΠ², сфокусированная Π½Π° слоТных Π·Π°Π΄Π°Ρ‡Π°Ρ… рассуТдСния с провСряСмыми Ρ€Π΅Π·ΡƒΠ»ΡŒΡ‚Π°Ρ‚Π°ΠΌΠΈ, Ρ‚Π°ΠΊΠΈΡ… ΠΊΠ°ΠΊ ΠΌΠ°Ρ‚Π΅ΠΌΠ°Ρ‚ΠΈΠΊΠ°, ΠΏΡ€ΠΎΠ³Ρ€Π°ΠΌΠΌΠΈΡ€ΠΎΠ²Π°Π½ΠΈΠ΅ ΠΈ STEM. ## ДоступныС ΠΊΠ²Π°Π½Ρ‚Ρ‹ | Π€Π°ΠΉΠ» | ΠšΠ²Π°Π½Ρ‚ΠΎΠ²Π°Π½ΠΈΠ΅ | Π Π°Π·ΠΌΠ΅Ρ€ | ОписаниС | | --- | --- | --- | --- | | **[VibeThinker-3B-f16.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-f16.gguf)** | FP16 | 6.18 Π“Π‘ | Π˜ΡΡ…ΠΎΠ΄Π½Π°Ρ Ρ‚ΠΎΡ‡Π½ΠΎΡΡ‚ΡŒ, максимальноС качСство. | | **[VibeThinker-3B-Q8_0.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-Q8_0.gguf)** | Q8_0 | 3.29 Π“Π‘ | ВысокоС качСство, рСкомСндуСтся для ΠΌΠΎΡ‰Π½Ρ‹Ρ… ПК. | | **[VibeThinker-3B-Q5_K_M.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-Q5_K_M.gguf)** | Q5_K_M | 2.22 Π“Π‘ | ΠžΡ‚Π»ΠΈΡ‡Π½Ρ‹ΠΉ баланс ΠΌΠ΅ΠΆΠ΄Ρƒ Ρ€Π°Π·ΠΌΠ΅Ρ€ΠΎΠΌ ΠΈ качСством. | | **[VibeThinker-3B-Q4_K_M.gguf](https://huggingface.co/KakTakOne/VibeThinker-3B-GGUF/blob/main/VibeThinker-3B-Q4_K_M.gguf)** | Q4_K_M | 1.93 Π“Π‘ | Π‘Π°ΠΌΡ‹ΠΉ быстрый ΠΈ Π»Π΅Π³ΠΊΠΈΠΉ ΠΊΠ²Π°Π½Ρ‚ (Π½Π°ΠΈΠ±ΠΎΠ»Π΅Π΅ популярный для повсСднСвного использования). | --- ## Π’Π²Π΅Π΄Π΅Π½ΠΈΠ΅ VibeThinker-3B ΠΏΡ€ΠΎΠ΄ΠΎΠ»ΠΆΠ°Π΅Ρ‚ Ρ€Π°Π·Π²ΠΈΡ‚ΠΈΠ΅ сСрии ΠΌΠΎΠ΄Π΅Π»Π΅ΠΉ рассуТдСния VibeThinker Π½Π° ΠΌΠ°ΡΡˆΡ‚Π°Π±Π΅ 3 ΠΌΠΈΠ»Π»ΠΈΠ°Ρ€Π΄ΠΎΠ² ΠΏΠ°Ρ€Π°ΠΌΠ΅Ρ‚Ρ€ΠΎΠ². Благодаря ΠΎΠΏΡ‚ΠΈΠΌΠΈΠ·Π°Ρ†ΠΈΠΈ ΠΏΠ°ΠΉΠΏΠ»Π°ΠΉΠ½Π° обучСния Spectrum-to-Signal Principle (SSP), модСль дСмонстрируСт Π²Ρ‹Π΄Π°ΡŽΡ‰ΠΈΠ΅ΡΡ Ρ€Π΅Π·ΡƒΠ»ΡŒΡ‚Π°Ρ‚Ρ‹ Π½Π° Π±Π΅Π½Ρ‡ΠΌΠ°Ρ€ΠΊΠ°Ρ… AIME, HMMT, IMO-AnswerBench, LiveCodeBench ΠΈ Π½Π΅Π΄Π°Π²Π½ΠΈΡ… контСстах LeetCode, ΠΏΡ€ΠΈΠ±Π»ΠΈΠΆΠ°ΡΡΡŒ ΠΏΠΎ качСству ΠΊ флагманским коммСрчСским модСлям рассуТдСния Π²Ρ€ΠΎΠ΄Π΅ Qwen3.6 Plus, Gemini 3 Pro, GLM-5 ΠΈ Kimi K2.5. ## ΠšΠ»ΡŽΡ‡Π΅Π²Ρ‹Π΅ ΠΏΠΎΠΊΠ°Π·Π°Ρ‚Π΅Π»ΠΈ ΠΏΡ€ΠΎΠΈΠ·Π²ΠΎΠ΄ΠΈΡ‚Π΅Π»ΡŒΠ½ΠΎΡΡ‚ΠΈ * πŸ“ МодСль Π½Π°Π±ΠΈΡ€Π°Π΅Ρ‚ **76.4** Π½Π° слоТном Π±Π΅Π½Ρ‡ΠΌΠ°Ρ€ΠΊΠ΅ IMO-AnswerBench (400 ΠΎΠ»ΠΈΠΌΠΏΠΈΠ°Π΄Π½Ρ‹Ρ… Π·Π°Π΄Π°Ρ‡ уровня IMO) с использованиСм всСго 3 ΠΌΠ»Ρ€Π΄ ΠΏΠ°Ρ€Π°ΠΌΠ΅Ρ‚Ρ€ΠΎΠ², ΠΈ ΡƒΠ»ΡƒΡ‡ΡˆΠ°Π΅Ρ‚ Ρ€Π΅Π·ΡƒΠ»ΡŒΡ‚Π°Ρ‚ Π΄ΠΎ **80.6** с ΠΏΡ€ΠΈΠΌΠ΅Π½Π΅Π½ΠΈΠ΅ΠΌ CLR (Claim-Level Reliability Assessment) Π½Π° этапС инфСрСнса. Π­Ρ‚ΠΎ сопоставимо с показатСлями Π³ΠΎΡ€Π°Π·Π΄ΠΎ Π±ΠΎΠ»Π΅Π΅ ΠΊΡ€ΡƒΠΏΠ½Ρ‹Ρ… ΠΌΠΎΠ΄Π΅Π»Π΅ΠΉ, Ρ‚Π°ΠΊΠΈΡ… ΠΊΠ°ΠΊ DeepSeek V3.2 (78.3, 671B), GLM-5 (82.5, 744B) ΠΈ Kimi K2.5 (81.8, 1T). * πŸ† На Π΅ΠΆΠ΅Π½Π΅Π΄Π΅Π»ΡŒΠ½Ρ‹Ρ… ΠΈ Π΄Π²ΡƒΡ…Π½Π΅Π΄Π΅Π»ΡŒΠ½Ρ‹Ρ… сорСвнованиях LeetCode (Python) Π·Π° ΠΏΠ΅Ρ€ΠΈΠΎΠ΄ с 25 апрСля ΠΏΠΎ 31 мая 2026 Π³ΠΎΠ΄Π° модСль ΡƒΡΠΏΠ΅ΡˆΠ½ΠΎ ΠΏΡ€ΠΎΡˆΠ»Π° **123 ΠΈΠ· 128** тСстов с ΠΏΠ΅Ρ€Π²ΠΎΠΉ ΠΏΠΎΠΏΡ‹Ρ‚ΠΊΠΈ (доля ΡƒΡΠΏΠ΅ΡˆΠ½Ρ‹Ρ… Ρ€Π΅ΡˆΠ΅Π½ΠΈΠΉ составляСт **96.1%**). ## Пайплайн обучСния ΠžΠ±ΡƒΡ‡Π΅Π½ΠΈΠ΅ VibeThinker-3B основано Π½Π° ΠΌΠ΅Ρ‚ΠΎΠ΄ΠΎΠ»ΠΎΠ³ΠΈΠΈ **Spectrum-to-Signal Principle (SSP)**: 1. **Curriculum SFT Π² Π΄Π²Π° этапа**: сначала общая кодовая ΠΈ матСматичСская Π±Π°Π·Π°, Π·Π°Ρ‚Π΅ΠΌ слоТныС рассуТдСния с Π΄Π»ΠΈΠ½Π½Ρ‹ΠΌ контСкстом. 2. **Multi-domain RL** с Π°Π»Π³ΠΎΡ€ΠΈΡ‚ΠΌΠΎΠΌ MaxEnt-Guided Policy Optimization (MGPO) Π² ΠΎΠΊΠ½Π΅ контСкста 64K. 3. **ΠžΡ„Π»Π°ΠΉΠ½ дистилляция Π½Π° сСбя (Self-Distillation)** для ΠΎΡ‚Π±ΠΎΡ€Π° Π»ΡƒΡ‡ΡˆΠΈΡ… Ρ‚Ρ€Π°Π΅ΠΊΡ‚ΠΎΡ€ΠΈΠΉ рассуТдСний. 4. **Instruct RL** для ΡƒΠ»ΡƒΡ‡ΡˆΠ΅Π½ΠΈΡ управляСмости ΠΈ форматирования ΠΎΡ‚Π²Π΅Ρ‚ΠΎΠ² ΠΏΠΎΠ΄ ΠΏΠΎΠ»ΡŒΠ·ΠΎΠ²Π°Ρ‚Π΅Π»Ρ. --- ## Как ΠΈΡΠΏΠΎΠ»ΡŒΠ·ΠΎΠ²Π°Ρ‚ΡŒ Π­Ρ‚ΠΈ Ρ„Π°ΠΉΠ»Ρ‹ GGUF ΠΌΠΎΠΆΠ½ΠΎ Π·Π°ΠΏΡƒΡΠΊΠ°Ρ‚ΡŒ Π² **LM Studio**, **Ollama**, **llama.cpp** ΠΈ Π΄Ρ€ΡƒΠ³ΠΈΡ… совмСстимых ΠΊΠ»ΠΈΠ΅Π½Ρ‚Π°Ρ…. ### LM Studio ΠŸΡ€ΠΎΡΡ‚ΠΎ Π²Π±Π΅ΠΉ Π² строку поиска `KakTakOne/VibeThinker-3B-GGUF` ΠΈ скачай Π½ΡƒΠΆΠ½Ρ‹ΠΉ ΠΊΠ²Π°Π½Ρ‚. ### Запуск Ρ‡Π΅Ρ€Π΅Π· консоль (llama.cpp) ```bash llama-cli -m VibeThinker-3B-Q4_K_M.gguf -p "2+2=" -n 128 ```
--- *Quantized by [KakTakOne](https://huggingface.co/KakTakOne) using `llama-quantize`.*