Punching Four Hundred Times Above Its Weight Class

What it does VibeThinker-1.5B is a 1.5B-parameter dense model trained to reason through competitive mathematics and coding problems. Its creators claim it rivals or exceeds much larger models—including the initial 671B-parameter DeepSeek R1—on benchmarks like AIME24, AIME25, and HMMT25. The weights and evaluation scripts are released under the MIT license.

The interesting bit The post-training recipe, dubbed the “Spectrum-to-Signal Principle,” first floods the model with diverse solution paths via two-stage distillation, then applies a maximum-entropy policy optimization to amplify the correct signal. The team claims the whole post-training bill was just $7,800—pocket change next to the six-figure budgets cited for DeepSeek R1 and MiniMax-M1.

Key highlights

• Claims to beat the initial DeepSeek R1 on AIME24 (80.3 vs. 79.8), AIME25 (74.4 vs. 70.0), and HMMT25 (50.4 vs. 41.7).
• Reportedly trained for roughly $7,800, a 30× to 60× cost reduction versus the $294K–$535K cited for DeepSeek R1 and MiniMax-M1.
• Uses “Two-Stage Diversity-Exploring Distillation” followed by “MaxEnt-Guided Policy Optimization (MGPO)” to squeeze reasoning from a small parameter budget.
• Built for long-form inference: the authors suggest generation contexts up to 40,960 tokens with temperature at 0.6 or 1.0.

Caveats

• Explicitly scoped to competitive math and coding; the authors do not recommend it for general open-ended chat.
• The README is heavy on benchmark victories but light on training data composition, hardware requirements, or architectural ablations.

Verdict A curiosity for researchers and hackers probing the lower bounds of reasoning model size, but unsuitable if you need a general-purpose conversational assistant.