Spaces:

---

TeichAI

/

README

like 1

Running

Was thinking just this! Will start with the Cascade model since it fits on my personal GPU, fingers crossed it works out of the box :)

Model was made with a Qwen3 base so it worked out of the box :)

TeichAI/Nemotron-Cascade-14B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill
TeichAI/Nemotron-Cascade-14B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill-GGUF

I think your distills may be the best open source models to date. Well done to the entire team!

I think your distills may be the best open source models to date. Well done to the entire team!

Thanks for the kind words, I think they are improvements to the base models for sure. Not sure if they are the best 😉

There are plenty of amazing teams out there that are doing a lot more than these relatively simple distillations, their work is amazing.

Either way, Nemotron Cascade 8B Thinking is up now as well:
TeichAI/Nemotron-Cascade-8B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill
TeichAI/Nemotron-Cascade-8B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill-GGUF

Please let me know which (if any other) models you would like me to distill into nemotron cascade next (i.e. claude 4.5 sonnet, deepseek v3.2, gpt 5.2, gemini 3 pro preview, etc)

They are quite impressive! I'll have to try the 8b!

Fwiw, I had a chance to try Nemotron-Cascade-14B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill and it seemed to give a truncated response. I have a collection of prompts I use as a general heuristic to gauge the quality of every LLM I test. Ex: "Write a production grade web application." I test with several languages but typically Ill use go lang as the suggested language. It was able to plan/think with similar output (though it takes this model longer to think) but the response itself seemed to be truncated compared with even the smaller Qwen3-4B-Thinking-2507-Claude-4.5-Opus-High-Reasoning-Distill.

I can post the responses I saw but maybe I just need to tune the hyper parameters a bit.

If you have a chance - https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 would be amazing!

it seemed to give a truncated response.

I will look into this, perhaps it is due to the models relatively low context length, but i find it hard to believe it's reaching it's limit on a single prompt. Not the slightest idea what could be causing it though. may roll back a couple hundred steps. If you get a chance to test out the 8B variant, please let me know if you have the same issue

If you have a chance - https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 would be amazing!

Waiting on my free colab credits for the month 😄

it seemed to give a truncated response.

After my first couple tests I see what you mean. Never experienced an issue like this before, give me some time to look into it more