LLaDA2.1-mini

LLaDA2.1-mini is the successor to LLaDA2.0-mini, the diffusion line that InclusionAI, the open-source arm of Ant Group, publishes alongside its conventional Ling models. The weights went up on 9 February 2026 under Apache 2.0. The architecture is unchanged from its predecessor: a mixture-of-experts diffusion backbone with 16.26 billion parameters in BF16, 256 routed experts of which 8 are used per token plus one shared, 20 layers, a 157,184-token vocabulary and a 32,768-token context. What changed is the training and the sampler, and it addresses the structural weakness of diffusion text generation. An autoregressive model commits to each token in order; a diffusion model unmasks a whole block at once, which is faster but leaves no way to fix a token that turns out wrong once neighbouring tokens are filled in. This version can edit already-unmasked tokens during later denoising passes — the makers call it error-correcting editable generation, and their paper is titled "Speeding Up Text Diffusion via Token Editing". The model ships in two sampling regimes rather than two sets of weights. Speed Mode averages 5.34 tokens per forward pass on the maker's benchmark suite against 2.60 for LLaDA2.0-mini — slightly more than double — while the average score drops from 63.39 to 62.07. Quality Mode averages 3.12 tokens per pass and 63.90, so it is both faster and marginally better than the older model. The speed-up is very uneven by task: on HumanEval+ Speed Mode reaches 12.32 tokens per pass against 5.16 for the predecessor, but on the Chinese knowledge test C-EVAL only 3.39 against 1.78. Code, with its predictable structure, is where parallel unmasking pays; open-ended knowledge questions are where it does not. One comparison in the publisher's own table is worth reading carefully. Ling-mini-2.0, the company's ordinary autoregressive model of the same size class, averages 64.72 — above both diffusion modes. On this evidence the diffusion line is not sold as the more capable technology but as the faster one, and the maker prints the numbers that say so. All benchmark figures here are the publisher's. The licence is stated as Apache 2.0 in both the card text and the metadata, but as with every other release from this publisher the repository ships no licence file at all — 17 files, none of them a licence. Downloads sit at about 101,000 in the last 30 days, roughly half of the 201,000 recorded for the older LLaDA2.0-mini in the same window.

▸Videos

No videos yet.

▸Reviews