Ant Group just open-sourced a 560B agent model under MIT, and the base checkpoint too 🧩

On 27 October, inclusionAI, the AI lab of Ant Group, published Ling-3.1-flash-1019 on Hugging Face. It's an upgraded version of Ling-3.1-flash with extra post-training aimed at reasoning and real-world agent tasks. What made me stop scrolling wasn't only the size. It was the licence: MIT, one of the most permissive licences there is, for a frontier-scale model. And two days earlier, the team also released Ling-3.1-flash-base, the raw pretrained checkpoint, under the same licence.

A big model that thinks small per token

Ling-3.1-flash-1019 has 560B parameters in total, but only about 25B are active for each token. It is a mixture-of-experts model with 512 routed experts, of which 8 (plus one shared expert) fire per token. The attention stack is the part I find most interesting: 63 layers of KDA, a linear-attention variant, combined with 9 Gated MLA layers, in a 7:1 ratio. Linear attention keeps the cost of long inputs under control, while the few full-attention layers help the model still find the exact detail it needs. The model card advertises a 1M-token context window.

Why I find this exciting

According to the model card, the extra post-training lifted its AA-Briefcase score from about 1400 to 1524 Elo and AutomationBench-AA from 62% to 68%, with gains on science tasks too (CritPt 18% to 26%, SciCode 54% to 57%). Overall it scores 44 on the Artificial Analysis Intelligence Index, up from 41 for its predecessor. Those are exactly the multi-step, tool-using workloads that agents live on.

But the part I value most is the base model. Releasing a base checkpoint means researchers and companies can do their own continued pretraining, domain adaptation and post-training, instead of only fine-tuning someone else's finished chat model. For anyone interested in distillation, a large MIT-licensed teacher with a matching base model is a very useful starting point. The team even ships day-0 support: a dedicated SGLang image with speculative decoding, plus vLLM instructions with tool-calling and reasoning parsers.

The honest caveats

This is not a homelab model. The reference SGLang setup runs on 8 H200 GPUs (or 4 GB300s), so most of us will meet it through an API or OpenRouter. The benchmark gains are the developer's own figures; Neural Brew notes that only the overall index was visible on Artificial Analysis when they checked. And while the card advertises 1M tokens of context, the shipped config.json sets max_position_embeddings to 262,144, so check your serving setup before throwing a whole codebase at it.

Sources

Open weights, an open base model and a permissive licence: that's the kind of release that lets the whole community build further. I'm curious what people will distil from it.

Reacties

Een reactie posten

Populaire posts van deze blog

Meta just shared the full recipe for a 30B agent model you can run at home 🧠

When Meta released Muse Glimmer in August, the pitch was simple: an open-weight, Apache 2.0 model built for AI agents that run on your own Mac or PC. What was missing was the "how". On 19 October, Meta Superintelligence Labs published the full technical report, and for anyone who cares about distillation and local agents, it is one of the most useful reads in a while. What the model is Muse Glimmer is a dense model of about 29.6 billion parameters, including a 1.8B perception encoder, so it reads text and images and writes text. It has a 128K+ context window and was trained on data from more than a hundred languages. The design goal was strict: everything has to fit in 32 GB of GPU memory, and a smaller 4-bit variant targets 24 GB cards. The parts I found most interesting Dense beat mixture-of-experts here. That surprised me. The team trained scaling ladders for a dense model and two MoE designs with the same memory footprint. Under a fixed memory ceiling and their compu...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

A 35B model on a Mac mini, using just 2.9 GiB of memory 🚀

There's something special about running a capable language model on hardware you own: no API bill, no data leaving the house, just a small box on your desk doing the thinking. A paper that's currently #2 on Hugging Face's trending list takes that idea further than I expected: it runs a 35-billion-parameter model on a Mac mini at about 20 tokens per second while using under 3 GiB of memory. The trick: leave the experts on the SSD The paper is called The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction , and the system is named Edge0 . It builds on a Mixture-of-Experts (MoE) model, Qwen3.6-35B-A3B. In an MoE, each token only uses a few "experts" per layer (here 4 out of 256 in each of 40 layers), so most of the weights sit idle at any given moment. Edge0 takes that literally. The expert weights (quantized to int4) stay on the SSD and are streamed in only when needed. The clever part is a small prerouter : a tiny trained he...
Meer lezen