What is VibeThinker-3B?
VibeThinker-3B is a 3.1B-parameter reasoning model from WeiboAI, fine-tuned from Qwen2.5-Coder-3B and aimed at one job: multi-step reasoning on problems whose answers can be checked, which in practice means competition math, competitive programming and STEM. WeiboAI published the weights on June 12, 2026 under the MIT license. The case for running it locally is the size: the 4-bit GGUF is a 1.93 GB file, so the model fits in about 4 GB of memory, which covers nearly any laptop.
| Specification | VibeThinker-3B |
|---|---|
| Total parameters | 3.1B (3,085,938,688) |
| Base model | Qwen2.5-Coder-3B |
| Modalities | Text in, text out (English) |
| Reasoning budget | Quick start defaults to 102,400 new tokens; 60K to 100K advised for the hardest problems |
| Post-training | Spectrum-to-Signal Principle: two-stage SFT, multi-domain RL, self-distillation, instruct RL |
| Tool calling | Not trained for it; the vendor advises against agent use |
| Recommended sampling | temperature 1.0, top_p 0.95 |
| Release date | June 12, 2026 |
| License | MIT |
The training recipe is why a 3B model shows up next to frontier names at all. WeiboAI's Spectrum-to-Signal Principle pipeline runs a two-stage curriculum SFT that deliberately preserves multiple valid solution paths, then reinforcement learning with verifiable rewards over math, code and STEM inside a 64K context window, then distills the strongest RL trajectories back into a single student model, and finishes with an instruct RL pass. The authors call the bet behind it the Parametric Compression-Coverage Hypothesis: verifiable reasoning compresses well into few parameters, broad world knowledge does not. They are open about both halves of that trade and recommend larger general-purpose models for open-domain tasks. The card also carries a plain warning: the model was not trained on tool-calling or agent data, so point it at LeetCode-style problems, not at function calling or autonomous coding agents.
VibeThinker-3B benchmarks
The numbers are WeiboAI's own, published on the model card. The headline is IMO-AnswerBench, a benchmark of 400 IMO-level problems, where the 3B lands in the range of models hundreds of times its size:
| Benchmark | VibeThinker-3B | DeepSeek V3.2 | GLM-5 | Kimi K2.5 |
|---|---|---|---|---|
IMO-AnswerBench Olympiad math | 76.4 | 78.3 | 82.5 | 81.8 |
IMO-AnswerBench with CLR Verified reasoning | 80.6 | 78.3 | 82.5 | 81.8 |
LeetCode contests Contest coding | 96.1% | - | - | - |
Read the first two rows with parameter counts in mind: DeepSeek V3.2 is 671B, GLM-5 is 744B and Kimi K2.5 is 1T, and at 3B the raw 76.4 still trails all three. Claim-Level Reliability Assessment, WeiboAI's test-time scaling strategy for answer-verifiable tasks, lifts the same benchmark to 80.6, which passes DeepSeek V3.2 but stays behind GLM-5 and Kimi K2.5. The LeetCode row is not a comparison: it is 123 of 128 first-attempt passes on unseen weekly and biweekly contests from April 25 to May 31, 2026, and WeiboAI publishes no competitor numbers next to it. The card also cites results on AIME, HMMT and LiveCodeBench without publishing a numeric table.
VibeThinker-3B hardware requirements
The system requirement to check is memory. The GGUF builds and file sizes below come from the community repo prithivMLmods/VibeThinker-3B-GGUF.
| Memory | Build to pick | File size |
|---|---|---|
| 4 GB | Q4_K_M | 1.93 GB |
| 6 GB | Q5_K_M | 2.22 GB |
| 8 GB | Q6_K | 2.54 GB |
| 12 GB | Q8_0 | 3.29 GB |
| 16 GB and up | BF16 | 6.18 GB |
Neighbouring files differ by a few hundred megabytes, so when two builds both fit, take the larger one. Leave headroom beyond the file itself: WeiboAI advises 60K to 100K token limits for the hardest problems, and a reasoning trace that long needs KV cache on top of the weights. If the format is new to you, start with what GGUF is.
How to run VibeThinker-3B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for VibeThinker-3B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
For the rest of the family, see every VibeThinker model you can run locally, or compare it with another small reasoning model, Qwen3-4B Thinking 2507.
VibeThinker-3B license
VibeThinker-3B is released under the MIT license. That permits commercial use, modification, fine-tuning and redistribution, with the single obligation of keeping the copyright and license notice, so you can build on the model and ship it in a product without a usage fee.