sirunchained/VibeThinker-1.5B-gguf

VibeThinker-1.5B GGUF

This is a quantized version of the WeiboAI/VibeThinker-1.5B model, converted to GGUF format for use with llama.cpp and compatible tools.

Quantization Details

The model was quantized using llama.cpp to the following formats:

• FP16: Converted from the original checkpoint as an intermediate step.
• Q4_K_M: Quantized using the Q4_K_M method.
• Q5_K_M: Quantized using the Q5_K_M method.
• Q6_K: Quantized using the Q6_K method.
• Q8_0: Quantized using the Q8_0 method.

Original Model Card Summary

• Model ID: WeiboAI/VibeThinker-1.5B
• Original Repository: https://huggingface.co/WeiboAI/VibeThinker-1.5B

Files Provided

• VibeThinker-1.5B-f16.gguf (FP16)
• VibeThinker-1.5B-q4_k_m.gguf (Q4_K_M)
• VibeThinker-1.5B-q5_k_m.gguf (Q5_K_M)
• VibeThinker-1.5B-q6_k.gguf (Q6_K)
• VibeThinker-1.5B-q8_0.gguf (Q8_0)

Q4_K_M Benchmark Results

The following results were obtained with the Q4_K_M quantization. Evaluation settings: temperature=0.6, top_p=0.95, max_gen_toks=40960, seed=1234. These settings match the official recommendation from the WeiboAI model card.

Benchmark	Metric	Q4_K_M Result	Sample Size	Original Result
custom_aime24	Accuracy	80.0% (±7.43%)	30	80.3
custom_hmmt25	Accuracy	46.67% (±9.26%)	30	50.4
custom_lcb_v6	Pass@1	41.98% (±4.33%)	131	51.1
custom_math500	Accuracy	96.80% (±0.79%)	500	95.0

Footnotes:

• The "Original Result" column refers to scores reported in the VibeThinker-1.5B technical report (Tiny Model, Big Logic, arXiv:2511.06221). MATH-500 (95.0) is from Table 1; the report evaluates it using average pass rate across 64 sampling trials.
• The custom_* task names refer to the official evaluation datasets loaded through a rewritten YAML configuration with a modified scorer. The datasets themselves are the standard benchmarks.
• The original model's benchmark scores were obtained with the same generation parameters recommended for reproduction: temperature=0.6, max_token_length=40960, top_p=0.95.

Key observations:

• AIME24: Q4_K_M (80.0%) vs. Original (80.3%) - essentially identical. 4-bit quantization preserves mathematical reasoning on this benchmark with negligible degradation.
• HMMT25: Q4_K_M (46.67%) vs. Original (50.4%) - a ~3.7-point drop, consistent with expected precision loss on a harder competition math benchmark.
• LiveCodeBench v6: Q4_K_M (41.98%) vs. Original (51.1%) - code generation is more sensitive to quantization, and the ~9-point gap reflects that.
• MATH-500: Q4_K_M (96.8%) actually exceeds the Original score (95.0). This benchmark has sufficient headroom that the quantized model retains full capability.

Original Model Performance

From the technical report:

Benchmark	Score
AIME24	80.3
AIME25	74.4
MATH-500	95.0
HMMT25	50.4
LiveCodeBench v5	55.9
LiveCodeBench v6	51.1

These scores place VibeThinker-1.5B on par with or above models 100×–600× larger, including DeepSeek R1 (671B) and Kimi K2 (1000B+).

Model tree for sirunchained/VibeThinker-1.5B-gguf

• Base model: Qwen/Qwen2.5-1.5B
• Quantized from: WeiboAI/VibeThinker-1.5B

---

Notes on the Results

The evaluation JSON files are labeled custom_aime24, custom_hmmt25, custom_lcb_v6, and custom_math500. These use the official benchmark datasets, run through a rewritten YAML configuration with a modified scorer (the default scorer would otherwise fail on these tasks). The generation parameters match the recommended inference settings from the original model card: temperature=0.6, max_token_length=40960, top_p=0.95.

Standard errors are included in the results table. The relatively large standard errors for the 30-sample benchmarks (AIME24, HMMT25) reflect the small sample size; MATH-500 is much more statistically stable with 500 problems.