MODEL · WEIBO AI · 3B DENSE
VibeThinker-3B
A 3B reasoning model that punches far above its size on VERIFIABLE reasoning — math, competitive coding, STEM. Weibo AI reports it reaching the range of much larger models (DeepSeek V3.2, GLM-5, Kimi K2.5) on IMO-AnswerBench (76.4, → 80.6 with a test-time verification strategy) despite only 3B params, via their Spectrum-to-Signal post-training. The thesis: verifiable reasoning is a parameter-dense, compressible capability where small models can reach near-frontier.
License: MIT · Context: Long — 60K–100K recommended for hard math · Released: June 12, 2026
The decision in five lines
- The call
- Consider — runnable locally, family reference
- Best for
- Local evaluation and family reference
- Runs on
- 23 hardware picks fit (cheapest: Intel Arc B580 12 GB · $249)
- Watch out
- Tool-calling, agent orchestration, or autonomous coding agents — the authors explicitly say it was NOT trained for those.
- Evidence
- Estimated
- 3B dense
- PARAMETERS
- SMALL REASONING MODEL
- TYPE
- Long
- CONTEXT
- ~2–3 GB
- VRAM AT Q4
Where we recommend this
This model isn’t currently in an active planner slot. See the runner notes below if you’re running it anyway.
The call
A 3B reasoning model that punches far above its size on VERIFIABLE reasoning — math, competitive coding, STEM. Weibo AI reports it reaching the range of much larger models (DeepSeek V3.2, GLM-5, Kimi K2.5) on IMO-AnswerBench (76.4, → 80.6 with a test-time verification strategy) despite only 3B params, via their Spectrum-to-Signal post-training. The thesis: verifiable reasoning is a parameter-dense, compressible capability where small models can reach near-frontier.
When not to use: Tool-calling, agent orchestration, or autonomous coding agents — the authors explicitly say it was NOT trained for those. Also weak on open-domain knowledge / general chat (small models cover facts poorly). Use it for math / contest-coding / STEM reasoning, not as a generalist.
Runner notes
Runs anywhere (3B) — Ollama / llama.cpp / transformers. Set a high max-token budget (60K–100K) for hard math; it thinks long. For agentic or tool work, use Qwen3-Coder-30B-A3B instead.
Hardware that fits
Every hardware pick whose memory fits this model at the quant we recommend. Sorted cheapest-first — the top row is your best-value fit. Click through for the full buyer’s guide.
- Intel Arc B580 12 GBPerfect · 3.9× 12 GB · $249–$299
- NVIDIA RTX 3060 12 GBPerfect · 3.9× 12 GB · $280–$400
- Minisforum UM890 ProPerfect · 7.8× 32 GB DDR5 (shared) · $463–$580 all-in
- RTX 5060 Ti 16 GBPerfect · 5.2× 16 GB · $560–$610
- AMD Radeon RX 9070 XTPerfect · 5.2× 16 GB · $649–$779
- Mac Mini M4 16 GBPerfect · 3.5× 16 GB unified · $799 (new floor) / $499–$599 (eBay/residuals)
- AMD Radeon RX 7900 XTXPerfect · 7.8× 24 GB · $810 used / ~$1,340 new
- NVIDIA RTX 3090 (used, single)Perfect · 7.8× 24 GB · $950–$1,200
- NVIDIA RTX 5070 TiPerfect · 5.2× 16 GB · $980–$1,300
- NVIDIA RTX 5080Perfect · 5.2× 16 GB · $999–$1,400
- MacBook Air M5 24 GBPerfect · 5.2× 24 GB unified · $1,499–$1,899
- Mac Mini M4 Pro 24 GBPerfect · 5.2× 24 GB unified · $1,599
- Dual RTX 3090 (used)Perfect · 15.6× 48 GB · $1,800–$2,500 all-in
- Framework Desktop (Ryzen AI Max+ 395)Perfect · 27.9× 128 GB unified · $1,999–$2,851
- NVIDIA RTX 4090Perfect · 7.8× 24 GB · $2,200–$2,800
- M5 Pro MacBook Pro 48 GBPerfect · 10.5× 48 GB unified · $2,999–$3,599
- NVIDIA RTX 5090Perfect · 10.4× 32 GB · $3,500–$4,300
- NVIDIA RTX A6000 (48 GB, used)Perfect · 15.6× 48 GB ECC · $3,500–$4,500
- Mac Studio M4 Max 64 GBPerfect · 13.9× 64 GB unified · $3,799
- NVIDIA DGX SparkPerfect · 27.9× 128 GB unified · $4,699
- M5 Max MacBook Pro 64 GBPerfect · 13.9× 64 GB unified · ~$5,199 (est.; June 25 2026 increase)
- Mac Studio M3 Ultra 96 GBPerfect · 20.9× 96 GB unified · $5,299
- Dual RTX 5090Perfect · 20.8× 64 GB (2×32) · $8,500–$10,500
Next step
Find-by-model — see what hardware runs this→