Code generation no longer has to be a token at a time. plaidq, a 0.7b continuous diffusion language…
code generation no longer has to be a token at a time. plaidq, a 0.7b continuous diffusion language model released september 10, writes code after one distilled denoising step, repurposing a pretrained autoregressive model as a bidirectional denoiser.
diffusion codegen is small and it is fast.
Context
The arXiv preprint 2609.04531, submitted 3 September 2026, presents PlaidQ as a 0.7B continuous diffusion language model for code generation that repurposes a pretrained autoregressive model as a bidirectional denoiser over continuous token embeddings. A 16-step student reaches 31.78 and 40.49 pass@10 on HumanEval and MBPP+, surpassing the same teacher sampled for 512 steps, and the single-step result is 7.07 pass@1 on HumanEval. The model card on Hugging Face shows a bidirectional Qwen3-0.6B trunk, 717.4M parameters, Apache 2.0 and an export timestamp of 4 September 2026.
The single-step claim is supported but modest: 7.07 pass@1 is low in absolute terms, and the paper's wording is functionally correct programs, not competitive quality. Fast was not quantified in the abstract read, and no latency claim is made here. The released September 10 date does not match the sources, which give 3 and 4 September, so it is unverified. It is a preprint. The paper compares at matched scale with discrete diffusion language models, whose own papers were not read. That diffusion codegen is small and fast is the author's thesis.
Watch next
- Independent reproduction and later paper versions with latency numbers.
Sources
- PlaidQ (arXiv 2609.04531)arxiv.org
- plaidq-0.7b (Hugging Face)huggingface.co
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 11:19 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →