Excellent, Fast Local Model (somewhat of a review)
This model is fast due to its 8B-A1B~ MoE+KDA architecture and very performant for its size. I've found that giving it web search and code execution (e.g. Python or JavaScript sandboxes) can elevate the model's abilities greatly and it feels like a bigger model in some areas while still being lightning fast. Pretty much frontier for its size.
I have been using it for basic research, general search queries, and generating simple snippets of code with great success. It has replaced my 7-9B dense on-device models and works well as a subagent to augment a larger model to speed things up.
Having Ling-3.0-tiny/flash be available under the MIT license makes it all the more usable as well. Though, I agree with #3 that it'd be really nice to have the base model for fine-tuning/plain text completion.
Thank you inclusionAI/Ant Group for this excellent release. I hope to see more models from you in the future!
thx, the pretrain base is uploading
Guys you dropped models very different in size (Ling-3.0-flash - 124B MoE, Ling-3.0-tiny - 8B MoE).
Thank you for them.
Can you please also drop something in between- like 30B Moe?
To pipe in, I think around a 30B~ MoE is good for more advanced local usage such as basic agentic coding or more advanced research since it balances intelligence with speed (and fits on local devices). Example models that work well on local devices in this range are gemma-4-26B-A4B-it (creative writing, general assistant) and Qwen3.6-35B-A3B (coding and agentic).
The main reason is that consumer devices (e.g. an "average" laptop or desktop upgraded with 16, 32, or 64 GB of RAM) can run such a MoE at speeds of 15-40 tokens/second on CPU/CPU+GPU/iGPU whereas they'd have a much harder time running a dense model - it'd work, yes, but would be very slow.
Medium MoE models like 100-300B are excellent for server deployments, but are tougher to run locally. Sure, you can get maybe a few tokens/second streaming from an NVMe SSD, but prefill will be slow and the whole thing will be too sluggish for serious long horizon work.
Bottom line is, a 30B MoE is just a nice size to have available in a model family. It'll run on a lot of modern home computers, be more powerful than < 20B options, and even though a few models of this size exist, it's good to have options. It doesn't need to be 30B, but around that ballpark would be nice (25-40B). In my opinion, I think Ling might be a good candidate for a 30B-ish model as it uses KDA, meaning context doesn't take up as much memory as other models while still being good, making it perfect for long horizon tasks.
@Reverger Very curious to ask why would 30B be a sweet spot?
I think the current trend is that 10B and below is a dividing line: models in that range will tend to run on mobile devices, while models around 30B will be more suitable for local machines. Anything larger, and you’ll probably need to consider running it in the cloud.
For local PCs, Qwen 35B and 27B are currently some of the strongest contenders. Honestly, it feels like they don’t really have much competition at the moment.
@RichardBian
Hi.
The quickest answer is their exceptional popularity and that they are fast to run.
While previous commenters laying out sound arguments, it would be just enough to say that most of people merely have already set their local inference engines fit for splitting RAM/VRAM namely for 30B-35B models (as long as most of AI companies publish models in this range) to try out LLMs quickly and without extra friction.
One can compare popularity of ~30B models vs that of large models:
https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF - 6.5 millions of downloads, 890 likes
https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF - 176k of downloads, 290 likes.
10B models are not so common to run on desktops/laptops.
