Astral-4B-Preview is a reasoning-centric language model from the Astral series, built on Qwen3-4b-thinking-2507 and fine-tuned on the nvidia/AceReason-1.1-SFT dataset to deliver configurable, step-by-step logical reasoning for research and development use. By including a "Reasoning-level" directive in the system prompt, users can control the model’s depth of reasoning—from direct answers to ultra-detailed reasoning traces—enabling nuanced, structured responses tailored to diverse problem-solving needs. As a preview release, Astral-4B-Preview is ideal for evaluating advanced reasoning capabilities with user-guided depth control in scientific and technical tasks.
Model Files
File Name
Quant Type
File Size
Astral-4B-Preview.BF16.gguf
BF16
8.05 GB
Astral-4B-Preview.F16.gguf
F16
8.05 GB
Astral-4B-Preview.F32.gguf
F32
16.1 GB
Astral-4B-Preview.Q2_K.gguf
Q2_K
1.67 GB
Astral-4B-Preview.Q3_K_L.gguf
Q3_K_L
2.24 GB
Astral-4B-Preview.Q3_K_M.gguf
Q3_K_M
2.08 GB
Astral-4B-Preview.Q3_K_S.gguf
Q3_K_S
1.89 GB
Astral-4B-Preview.Q4_K_M.gguf
Q4_K_M
2.5 GB
Astral-4B-Preview.Q4_K_S.gguf
Q4_K_S
2.38 GB
Astral-4B-Preview.Q5_K_M.gguf
Q5_K_M
2.89 GB
Astral-4B-Preview.Q5_K_S.gguf
Q5_K_S
2.82 GB
Astral-4B-Preview.Q6_K.gguf
Q6_K
3.31 GB
Astral-4B-Preview.Q8_0.gguf
Q8_0
4.28 GB
Quants Usage
(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)
Here is a handy graph by ikawrakow comparing some lower-quality quant
types (lower is better):
