Seed 2.0 Lite Guide: Two Builds, One Name, a Sensible Default

Seed 2.0 Lite is the middle tier of ByteDance's Seed 2.0 family, and the tier ByteDance clearly wants most production traffic to land on. The catch is that "Seed 2.0 Lite" names two different builds, and the published numbers move in both directions between them. Before you compare it with anything, it helps to know which one you are calling.

Two builds behind one name

The Seed 2.0 family launched in February 2026 with Pro, Lite and Mini as its general-purpose models; a separate Code model reached OpenRouter later. ByteDance's official Seed 2.0 page now tracks two Lite versions: Lite (0215), the launch build, and Lite (0428), an end-of-April upgrade. ByteDance describes 0428 as the first omni-modal understanding model in the Seed line, adding native audio alongside video, image and text, with upgraded agent, coding and GUI capabilities.

OpenRouter added bytedance-seed/seed-2.0-lite on 10 March 2026 and lists text, image and video input, but not audio. It does not state which build it currently serves. If audio or the April GUI gains matter to you, check with your provider rather than assuming you have 0428. An undated name is a moving target: the build behind it can change without the identifier changing.

What changed between builds

The official page puts both builds next to Pro (0215). A selection:

• GPQA Diamond — Lite 0428 88.4%, Lite 0215 85.1%, Pro 88.9%
• SWE-Bench Pro — 46.6%, 46.0%, 46.9%
• SWE Multilingual — 66.6%, 64.4%, 71.7%
• Terminal Bench 2.0 — 43.3%, 45.0%, 55.8%
• BrowseComp — 64.0%, 72.1%, 77.3%
• HLE, text only, no tools — 25.7%, 28.2%, 32.4%
• OSWorld-Verified — 64.4% for Lite 0428; not listed for the others

The April build gained on knowledge-style reasoning and visual tasks, and on video benchmarks such as OmniVideoBench, where it moved from 44.5 to 61.7. It lost ground on BrowseComp, Terminal Bench and HLE. That is a normal pattern for a rebuild that adds modalities: capability shifts around rather than rising everywhere. If your workload is web research or terminal agents, the older build may actually suit you better, which is another reason to know which one you are getting.

These are ByteDance's own numbers on ByteDance's own harnesses. They are useful for comparing builds of the same model, less useful for comparing against other vendors. Vendor-claimed vs independent benchmarks explains the gap.

How close it gets to Pro

On SWE-Bench Pro and GPQA Diamond, Lite is within a point of Pro. The gap opens on long agentic tasks: Terminal Bench 2.0 shows Pro more than ten points ahead, and BrowseComp shows Pro ahead of both Lite builds. ByteDance positions Pro for long-chain reasoning and robustness in complex workflows, and the numbers back that up.

The practical reading: for single-turn coding, Q&A and visual understanding, Lite gives you most of Pro. For agents that run many steps with tools, Pro is measurably stronger, and the difference compounds over a long session.

What it costs

OpenRouter lists Seed 2.0 Lite at $0.25 per million input tokens and $2.00 per million output at the time of writing, rising to $0.50 and $4.00 for prompts over 128K tokens. Those are exactly the rates OpenRouter charges for Seed 1.6, which has a listed retirement date of 11 November 2026. If you are on Seed 1.6, Lite is the like-for-like replacement at no change in price.

The endpoint supports tool calling, JSON mode, structured outputs and a reasoning effort control, with 262,144 tokens of context and up to 131,072 tokens of output.

Lite, Mini or Code

The family has three neighbours, and the choice is mostly about volume and task shape:

• Choose Lite as the default for mixed production work: assistants, document and image understanding, moderate agent loops. It is the tier ByteDance designed to balance quality against response speed.
• Choose Seed 2.0 Mini when volume dominates. It costs $0.10 input and $0.40 output, so Lite is 2.5 times the input price and five times the output price. On high-concurrency classification, extraction or batch generation, that gap is the whole budget.
• Choose Seed 2.0 Code only if your own tests show it solving coding tasks in fewer turns. It costs more than Lite, and the official Seed 2.0 page lists no benchmark scores for it.

Setting reasoning effort

Lite exposes reasoning effort rather than an adaptive on/off switch. Start at a low setting for routine requests and raise it per route where you can measure a quality gain. Higher effort spends output tokens, and output is eight times the input price on this model, so effort is the main lever on what Lite actually costs you.

Takeaway

1. Find out which Lite build your provider serves.
2. Use it as the default tier, with Mini under it for volume and Pro over it for long agent loops.
3. Set reasoning effort explicitly per route and measure output tokens, not just quality.
4. If you are migrating from Seed 1.6, run a sample of real prompts through Lite before 11 November; the price is unchanged, so the only question is quality.