Seed 2.0 Mini Guide: Throughput First, Reasoning on a Dial

Small tiers are usually judged against their big siblings and found wanting, which misses the point. Seed 2.0 Mini is not trying to be a cheaper Pro. ByteDance built it for inference throughput and deployment density: lots of concurrent requests, batch generation, and latency budgets where a large model would never fit. The useful question is not how much worse it is, but on which tasks the gap stops mattering.

The official data answers part of that question unusually clearly, because the only benchmarks ByteDance published for Mini are video benchmarks, and they split neatly into tasks where Mini keeps up and tasks where it does not.

What ByteDance built it for

Mini launched with the Seed 2.0 family on 14 February 2026, alongside Pro, Lite and a Code model. ByteDance's launch material positions it for low-latency, high-concurrency and cost-sensitive scenarios; the official Seed 2.0 page says it is optimised for inference throughput and deployment density, for high-concurrency and batch generation work.

OpenRouter added it on 26 February 2026 and describes it as delivering performance comparable to Seed 1.6, with a 256K context, four reasoning effort modes and multimodal understanding. The listing shows 262,144 tokens of context, up to 131,072 tokens of output, text, image and video input, and support for tool calling, JSON mode and structured outputs. The weights are not open.

What the official numbers show

The Seed 2.0 page lists Mini (0215) only in its video table. Set against Lite (0215) and Pro (0215):

• VideoMME — Mini 81.2, Lite 87.7, Pro 89.5
• VideoMMMU — 80.6, 84.1, 86.9
• LongVideoBench — 74.8, 77.3, 80.3
• TVBench — 70.5, 71.5, 75.0
• VideoReasonBench — 40.5, 64.2, 77.8

The pattern is the useful part. On perception and comprehension benchmarks, Mini sits a few points behind Lite. On VideoReasonBench, which asks the model to reason over what it has seen, it falls more than twenty points behind Lite and over thirty behind Pro.

That is a reasonable rule of thumb beyond video, though it is an inference rather than a published result: Mini is close to its siblings on recognising and extracting, and much weaker on multi-step reasoning. Route accordingly. These are vendor-run figures, and ByteDance published no Mini numbers for coding, search or knowledge benchmarks, so test those yourself.

Four effort levels

OpenRouter lists four reasoning effort modes for Mini: minimal, low, medium and high. That gives you a cost dial on a single endpoint:

• Minimal for classification, routing decisions, tagging and extraction, where reasoning tokens are pure overhead.
• Low for short transformations and structured output that needs a little planning.
• Medium or high only on routes where you have measured a gain. Given the VideoReasonBench gap, a hard reasoning task is often better sent to a larger model than run at high effort on Mini.

Setting effort explicitly matters more on a small model than a large one, because the point of choosing it was cost and latency, and high effort spends both.

The cost case, worked through

At the time of writing OpenRouter lists Mini at $0.10 per million input tokens and $0.40 per million output, rising to $0.20 and $0.80 above 128K tokens of prompt. Seed 2.0 Lite lists at $0.25 and $2.00.

Take a pipeline that processes 10M input tokens and 2M output tokens a day. On Mini that is $1.00 plus $0.80, so $1.80 a day. On Lite it is $2.50 plus $4.00, so $6.50 a day. The output side drives most of the difference, which is why Mini wins hardest on work that writes a lot: batch generation, bulk summaries, synthetic data.

The saving only counts if quality holds. A cheap model that needs a retry on one request in five, or a human correction on one in twenty, can cost more than the tier above it. When a cheap model is enough covers how to measure that before committing.

Mini as a replacement for Seed 1.6 and Flash

Two older endpoints, Seed 1.6 Flash and Seed 1.6, have a listed OpenRouter expiration date of 11 November 2026. Mini is the obvious landing spot for both. Against Flash it costs about a third more per token, with the same context size and video input, plus finer effort control. Against Seed 1.6 it is much cheaper, and OpenRouter describes the performance as comparable. If you need exactly the old Seed 1.6 price point with a newer model, Lite matches it; if you want to cut cost, Mini is the move.

A routing pattern that suits it

Mini works best as the first tier in a router rather than the only model. Send everything to Mini at minimal effort, have it classify the request, and escalate the hard cases to Lite or Pro. Most traffic in most products is easy, and the expensive model then only sees the share that needs it. Model routing and fallbacks walks through how to build that without a fragile rules engine.

The decision rule is short: pick Mini over Lite when volume is high, answers are mostly recognition or extraction, and latency matters. Pick Lite when the task needs reasoning across several steps, or when a wrong answer costs more than the token saving.