Interesting new models

[rofl0r]

I figure it makes sense to have a generic discussion topic for interesting new models rather than one per model, to keep notification spam minimal.

I discovered Nanbeige 4.3 a few days ago, which has 33B params and benchmark claims suggest it should be considerably stronger than Qwen 3.8 27b, the current open source champion for discrete GPU owners. It even ships QAT quants OOTB. Did anyone test it already, and can share some insights/experiences ? It appears its release went almost unnoticed; there are no results on https://llm-bench.io so far. Nor was it tested on https://artificialanalysis.ai/ .

[rofl0r]

https://huggingface.co/blog/Diffuspace/clover-35b-a3b-preview Qwen-3.6 35b-a3b turned into a diffusion model, supposedly better generation speed and higher benchmark scores.

[rofl0r]

https://huggingface.co/GrootTeam/Groot-19B-A3B-261023 a down-scaled version of Qwen 3.6 35b-a3b to 19b total params, supposedly beating the parent in 12/16 benchmarks.

[rofl0r]

i tested this one, it's got quite crippled world knowledge and even goes into endless reasoning loops when asked to translate english text into a language that has been removed from its weights, but it showed strong agentic and coding skills, and successfully fixed an issue in llama.cpp that has been annoying me ( ggml-org/llama.cpp#29163 ) . speed is identical to its parent (only active params seem to influence it) and it's coding skills are roughly on the level of qwen 3.6's skills, so it's only interesting for someone with a very limited RAM budget, as other qwen 3.6 forks such as Ornith 1.5 and Tiel Coder perform considerably better (almost qwen 3.8 27b level) and preserve their world knowledge. it could be an interesting base to do ornith 1.5-style RL on it and a few tweaks on the weights to remove the looping on removed knowledge wouldn't hurt either. doing so could result in a very competent model that runs fast and comfortably in 16GB RAM with plenty room to spare for KV cache and the OS, while running at 10-20 t/s even on CPU only.

[rofl0r]

https://huggingface.co/InternScience/Sage-A1.5 another qwen 3.6 35b-a3b fork with supposedly even better scores than ornith 1.5. llm-bench runs are too few to draw a verdict, but they're definitely in the ornith 1.5 ballpark so far: https://llm-bench.io/benchmarks?search=sage