How about trying merge the AlexWortega/SIQ-1-35B models?
CoT in this model seems to show high efficiency with less tokens. I think it would be a good idea to try merging.
"The car wash is 50 meters away from home. I need to wash my car, should I walk or take a car? "
When I asked the model the question: tepirale/Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4 and SIQ both came to the conclusion that logically the car should be driven, but the SIQ produced an answer that was around 1000 tokens, and the V4 model uses almost 3000 tokens.
What is surprising is that your first model (tepirale/Ornith-Agents-A1-3.7-35B-A3B-dare_ties) and the InternScience/Agents-A1 model came to the conclusion that you should walk the same way on this question.
I'm very happy that I can see progress.
Yes, it really surprised me.
MODEL="${MODEL:-tepirale/Ornith-Agents-A1-3.6-35B-A3B-dare_ties}"
He indirectly called me lazy. HAHAHA
yes! The original A1 model and Base Qwen model(3.6 35b a3b) couldn't solve those simple logical questions, but version4 of your models answered this correctly.
Just like the picture!
I tested the Ornith-Agents-A1-3.7-35B-A3B-dare_ties model by passing it an image to extract all its text and then generating Python code to recreate the image with the text embedded. I was pleasantly surprised by the result, as I had only seen this capability in Claude and ChatGPT until now. I used it within a harness developed by Fable 5, along with subagents, a terminal, and more than 20 other tools, and it worked perfectly. It improves significantly with a harness.
harness agent is more efficient than hermes agent?
and why aren’t you used v4 model?
I asked Fable 5 to create a harness for the 'tepirale/Ornith-Agents-A1-3.6-35B-A3B-dare_ties' model, utilizing tools and sub-agents that call the model itself (via vllm serve); when it makes a mistake, it uses the sub-agents to write a prompt and recover from the error.
Interestingly, version 4 is the one that makes the fewest mistakes. However, I use version 1 on purpose because I know it's more prone to errors; this way, I can identify its weaknesses and keep them in mind when using version 4. That way, if version 4 makes a mistake that I've already detected in version 1, I'll know how to apply the appropriate patch or solution.
For now, I'm testing "Harness Agent," and it has worked for me.
Wow... I didn't realize there was such a deep strategy behind it. Definitely beyond my pay grade lol. I'll be cheering you on!
model 27b
tepirale/Qwen3.5-27B-ThinkSonic-Coder-dare_ties-MTP-GGUF
model: Q6_K.gguf
GPU RTX 3090 24GB VRAM
Thinking
temperature=0.6, top_p=0.95, max_tokens=38192,
top_k=65, min_p=0.0, repetition_penalty=1.05,
presence_penalty=0.0, frequency_penalty=0.0,
tok out: 389
No thinking:
thinking=False, max_tokens=38192, temperature=1,top_k=20,
tok out: 149
That's wonderful!!
Unlike the MoE model, the output token has decreased surprisingly much.
But I can't use that because apple m1 max is too slow in dense model...
I’m really looking forward to seeing your MoE model tuning as well!
Thank you so much for your interest, I honestly never thought anyone would use the MOE model.
AND...
WoooOOOw, you're the most downloaded MLX! I'd like to know your opinion on how it performs on that device since I don't have an Apple product. Positive and negative aspects model, if possible?
Thanks a lot!
You've made a great model, and I'd be happy to contribute to it. In fact, you're the one who got me interested in merge models!
The main advantage of Apple devices is that they can run large models at a relatively low cost. However, they still lack absolute performance compared to Nvidia—except for the recently released M5 chip, which has drastically improved prefill performance, though I think decode performance is still lower than Nvidia's.
That’s exactly why MoE models are so important for Apple hardware. Despite having a large total parameter count, an MoE model only activates around 3B parameters at a time. This provides a huge advantage in both prefill and decode speeds.
It's very gratifying to know I can contribute to the open-source LLM model community. Mergekit is an exceptional tool for those of us who don't have a GPU or powerful infrastructure at home. In my case, I rely heavily on renting GPUs by the hour to test my experiments and on the excellent labs that train and upload their models to Hugging Face. Thanks to this ecosystem, independent developers can merge LLM models and continue innovating together. I also didn't know that these MOE models were important in the Apple community.
Thank you very much for your contribution.
What I've noticed with these MOE models on Nvidia GPUs is their great speed. For example, with this model, I've achieved between 250 and 290 tokens/s with MTP enabled and between 160 and 200 tokens/s with it disabled. I've also observed how its capabilities have increased with these parameters.
Initially, I doubted my idea for improving a section of DARE-TIES would work, as my previous tests only yielded minimal improvements in the evaluation. However, thanks to the 'Fable' model, I was able to optimize my code draft. By applying this DARE-TIES technique, the metrics increased significantly and directly, without the need for retraining.