Seed-Coder

What it is

Seed-Coder is a family of open code models released by ByteDance Seed on 8 May 2025; the project had been known before that as Doubao-Coder.^[1] It ships in three variants — Base, Instruct and Reasoning — and all three have 8 billion parameters.^[1] The release is under the MIT license and comes with a technical report on arXiv numbered 2506.03524.^[1] The repository says all the Seed-Coder models are publicly available in a ByteDance-Seed collection on Hugging Face.^[1]

The data is the story

Descriptions of code models often start with architecture. Seed-Coder's authors start with the training data. The family is built on what the project calls "model-centric" data curation: large language models, rather than hand-written rules, do the filtering of code pre-training data drawn from GitHub, commits and code-related web pages.^[1] The technical report carries the same idea in its title, "Seed-Coder: Let the Code Model Curate Data for Itself".^[2]

Filtering matters because scraped code is wildly uneven. A repository may hold a polished library, a half-finished experiment, a generated file nobody reads, or a snippet quoted inside a web page without its surrounding context. Training on all of it teaches a model bad habits as readily as good ones. A rule-based filter can only remove what someone thought to write a rule for — file suffixes, licence headers, minimum length. A model-based filter can, in principle, judge sample by sample, catching cases nobody anticipated when the rules were written. That is the trade the project makes: put a language model in the place of the rule book, and let it decide what the code model gets to learn from.^[1]

Three variants, two context lengths

The Base and Instruct models use a 32K context window, while the Reasoning model uses 64K and was trained with reinforcement learning.^[1] A bf16 build of the Reasoning model was also published.^[1] The split follows a familiar division of labour: the Base model holds pre-trained knowledge of code, the Instruct model is tuned to follow a user's request directly, and the Reasoning model is trained to work through a problem before answering. The larger context on the Reasoning model gives it more room to hold a problem and its intermediate steps at once.

Reported results, and one correction

ByteDance Seed states that the models reach state-of-the-art results among open models at the 8B scale on code generation, completion, editing, reasoning and software-engineering tasks, and that they surpass some much larger models.^[1] Comparisons of this kind depend on which benchmarks are chosen and how they are run, which makes the second half of the story worth noting. On 14 May 2025 the repository published a note saying that it had found an inconsistent setting in its BigCodeBench evaluation results.^[1] Publishing that kind of correction openly is a small but useful signal: benchmark tables in a model release are a claim, and claims can be checked.

The paper

The technical report, "Seed-Coder: Let the Code Model Curate Data for Itself", was submitted to arXiv on 4 June 2025 and revised one day later, on 5 June 2025.^[2] It appeared about a month after the models themselves, so the weights and the Hugging Face collection were available before the full write-up was.

Where it sits

Seed-Coder is an 8B open-weight family under a permissive licence, aimed at the everyday work of writing, completing and repairing code.^[1] Its distinguishing claim is less about size than about method: the same kind of model being trained is also used to decide what the training data should be.^[1] For readers watching how code models are built, that is the part worth following — the argument that data curation is a modelling problem, not a cleaning script.