Introducing the v0 composite model family

A different way to build a coding model

The v0 model family — v0-1.5-md, v0-1.5-lg, and v0-1.0-md — pairs a general-purpose base LLM with specialized components layered around it. The composite design combines retrieval-augmented generation (RAG) for current documentation, a frontier LLM for reasoning, and a custom streaming post-processor that fixes errors as code is generated. The benefit of this split is that Vercel can swap in newer base models as they ship without reworking the entire pipeline.

Why not just use a frontier or open model?

Off-the-shelf models have two problems for a product like v0, which specializes in building full-stack web applications: their knowledge of fast-moving frameworks like React and Next.js goes stale quickly, and frontier labs have little incentive to optimize for niche tasks such as automatic error correction or rapid code editing. Open-source fine-tuning offers more control, but proprietary models still beat open models on v0's core tasks by a wide margin, particularly multimodal code generation.

The composite architecture decouples these concerns. Instead of forcing one model to handle retrieval, reasoning, and error correction, v0 pairs a strong base model with dedicated data retrieval, a fast Quick Edit pipeline, and a custom AutoFix model. When you use v0 via the API, you get the entire assembly, not just the base model.

Inside the pipeline

Before a model call, v0 pre-processes the request in several steps. The system prompt defines v0's response format and capabilities, recent chat messages are added for continuity, and older messages are summarized to keep the context window lean. Additional context is retrieved from v0's own dataset based on the query — pulling from documentation, UI examples, uploaded project sources, and internal Vercel knowledge.

For new generations or large changes, the base model — a high-capability frontier model chosen per the v0 tier — does the work. For narrow edits like updating text, fixing syntax, or reordering components, requests are routed to a faster Quick Edit model. The architecture kept the rest of the stack stable when base models changed: v0-1.0-md uses Anthropic's Sonnet 3.7, while v0-1.5-md runs on Sonnet 4.

As the base model streams output, a custom AutoFix model continuously inspects the stream for errors, inconsistencies, and policy violations, applying mid-stream fixes. After streaming finishes, a final pass catches remaining issues that couldn't be detected mid-stream, and a linter resolves style inconsistencies.

Benchmark picture

Vercel measures how often each model produces error-free code on eval sets built from common web development tasks. On these benchmarks, v0 models outperform their base model counterparts by a substantial margin.

Note: v0-1.5-lg is a larger model than v0-1.5-md but can make more mistakes — a familiar scaling tradeoff.

Error rates are similar across v0-1.5-md and v0-1.5-lg, but the larger model handles hyper-specialized domains better — physics engines in three.js, for instance — and is stronger at multi-step tasks like database migrations. It also supports a much larger context window.

Training a custom fixer

Every language model has quirks: some over-format with markdown, others misplace files or introduce subtle bugs. Vercal tracks these patterns via a broad eval suite and user feedback from v0.dev, targeting categories where output consistently needs correction.

The resulting pipeline combines deterministic rules with AI-based corrections, built into a custom AutoFix model called vercel-autofixer-01, trained with Fireworks AI using reinforcement fine-tuning (RFT). Multiple iterations reduced error rates across the tracked categories.

Each line above tracks a different error type being minimized during training.

Designed for speed while reducing error rates, vercel-autofixer-01 performs at par with gpt-4o-mini and gemini-2.5-flash on Vercel's error incidence evals while running 10 to 40 times faster.

The v0 model family is currently available through the API and v0.dev, and can be used in editors or custom workflows — including automated code migrations. New model classes are expected in the coming months.