inclusionAI has released the Ling-3.0 model family on Hugging Face under a plain MIT license, with an architecture that mixes cheap linear-attention blocks and expensive full-attention blocks in a fixed three-to-one rhythm. The smallest variant, Ling-3.0-tiny, holds 7.9 billion total parameters but activates only 1.3 billion per token across 128 routed experts. Within hours of release, practitioners had filled the model's discussion pages with work on llama.cpp support, chat template fixes and independent evaluation results.
Key facts
- Ling-3.0-tiny: 7.9B total parameters, 1.3B active, 128 routed experts, a 3-to-1 linear-to-full attention ratio.
- Licensed MIT, with native BF16, FP8 and INT4 support declared on the model card.
- Released alongside Ling-3.0-flash, whose discussion page filled with practitioner debugging within hours.
- Published by inclusionAI on Hugging Face.
Two things about this release are more interesting than the benchmark table, and neither of them is a score.
The first is the architecture becoming ordinary. Full attention -- where every token in the sequence looks at every other token -- is what makes transformers work and also what makes long contexts expensive, because the cost grows with the square of the length. Linear attention variants approximate the same operation at cost that grows in proportion to length rather than its square. They are cheaper and somewhat less expressive, and for several years they lived in research papers rather than in models people actually download.
Ling-3.0 interleaves them at a fixed ratio: three linear-attention blocks, then one full-attention block, repeating. The intuition is that most of the work of reading a sequence does not need every-token-to-every-token comparison; you need that occasionally, for the hard integration steps, and cheap approximate mixing the rest of the time.
An analogy: reading a long report by skimming three sections and carefully rereading the fourth, on a fixed rotation. You lose a little. You finish in a fraction of the time.
Layered on top is a mixture of experts design. The model contains 128 separate expert sub-networks, and a router sends each token to only a few of them. The consequence is the 7.9-billion-versus-1.3-billion split: the model knows as much as a 7.9 billion parameter model, and costs roughly what a 1.3 billion parameter model costs to run per token. The card also declares native BF16, FP8 and INT4 support, meaning the quantized variants are a supported path rather than a community afterthought.
The second interesting thing is the license. MIT is a plain, permissive software license -- do what you want, keep the notice. That is meaningfully different from the custom community licenses attached to several prominent open-weight releases, which add acceptable-use clauses, attribution requirements or revenue thresholds. Licensing has quietly become one of the sharper differentiators between open-weight releases; we covered a version of this when Qwen3.8-27B shared its predecessor's bones but not its contract.
The third thing worth noting is not in the release at all -- it is what happened after. Within hours, the model's Hugging Face discussion tabs filled with practitioners working on llama.cpp support, fixing chat templates and posting independent evaluation results. That public, vendor-hosted, same-day triage has become the real health indicator for an open-weight release, more informative than any number on the card.
The same process is visible, more painfully, on a competing family released the same week. The Ornith 1.0 line -- also MIT, also derived from Qwen 3.5, published at 9B, 35B and 397B -- has a discussion thread in which users worked out that multi-token prediction tensors appear to be missing from the shipped checkpoint. That turns "the speedup feature is slow" from a model-quality complaint into a packaging problem with a fix. Nobody had to file a support ticket and wait.
The honest caveat is that a permissive license on weights is not the same thing as an open model. Training data and training code remain undisclosed for both families, which means nobody outside the labs can audit what went in, reproduce the result, or check for contamination. And 128 experts with 1.3 billion active is excellent on paper and entirely dependent on a serving stack that routes efficiently -- expert routing that thrashes across devices can erase the theoretical saving completely. The independent evaluations arriving over the next week are the ones that will settle it.
Related reading on this site: open-weight models, mixture of experts, linear attention, and why LLM inference is memory bound.
Originally published on Ground Truth, where every claim is checked against the primary source.
