@artyomita: A 9B student beats DeepSeek V4 Pro on a math test. It is free, and the Q4_K_M file is 5.78 GB. Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B is a Qwen3.5-9B fine-tune (Apache-2.0), trained on about 250,000 math and STEM answers written by DeepSeek-V4-Pro. On the model card's GSM8K table it scores 94.5 against 92.6 for DeepSeek-V4-Pro. On an MMLU-Pro sample it writes 2,858 reasoning tokens per question, the base Qwen3.5-9B 4,425. A community version called heretic, which cuts the refusal direction out of layers 9 to 24, was first on Hugging Face trending on Nov 12: refusals 3/100 instead of 100/100. Save this. Which model should I decode next? Tell me in the comments. Code: github.com/ArtyomITA Music: "Late Night Radio" by Kevin MacLeod (incompetech.com), CC BY 4.0 #ai #llm #deepseek #qwen
adriandj3
Region: IT
Friday 13 November 2026 20:30:00 GMT
Music
Download
Comments
Thanatos One :
can running normal in my RTX 3050M 4GB ?
2026-11-15 11:07:12
0
TXI :
what if it gets abstract prompts avoiding to my logic it would preform worse than the larger model which was used to train it
2026-11-14 01:30:08
1
intelligence :
Why people thinking math is AI? Try to make something useful like coding or image generation or video generation in that size and in that score. Maybe I don’t know if there is better use case of this model.
2026-11-15 06:37:05
0
To see more videos from user @artyomita, please go to the Tikwm
homepage.