Smallest Seed 2.0 tier: 262K-token context, image and video input, and four reasoning-effort levels for high-concurrency workloads.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Vendor-Reported Math and Coding Scores Hold Up for the Smallest Seed 2.0 Tier
Technical Specifications
What the model supports
Comparison
Seed-2.0-Mini or Seed-2.0-Lite: Where Does the Quality Gap Matter?
Both are Seed 2.0 models with a 262,144-token context window, text, image and video input, and tool calling, so the choice is mostly quality versus cost and speed. In ByteDance's model card, Lite leads on AIME 2025 (93.0 vs 87.0), Codeforces (2233 vs 1644), SWE-Bench Verified (73.5 vs 67.9) and HLE (28.2 vs 13.3). Mini has materially lower output pricing and, in a point-in-time OpenRouter measurement, a higher median throughput (131 vs 54 tokens per second).
Choose Seed-2.0-Mini when volume, latency and per-request cost dominate: classification, extraction, moderation, routing and multimodal triage. Choose Seed-2.0-Lite when answer quality drives value: harder coding, deeper reasoning, or tasks where Mini's weaker factuality (FactScore 50.4 vs 62.4) creates review overhead. A practical pattern is Mini as the default tier and Lite as the escalation tier, validated on your own evaluation set.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
High-volume classification and extraction
ByteDance positions Mini for latency-sensitive, high-concurrency, cost-sensitive scenarios, and it supports structured outputs and tool calling. In the model card its instruction-following scores include 91.2 on COLLIE and 61.1 on MultiChallenge, the latter above GPT-5 mini High (59.0). It fits intent tagging, field extraction and ticket routing at minimal reasoning effort. Validate outputs against a schema, since factual recall is weak.
Image and video triage
Mini accepts image and video input and returns text, which suits content moderation, visual classification and caption-style tagging. ByteDance reports 81.2 on VideoMME for the model, below Lite and Pro. It works as a first-pass labeler where volume matters more than fine-grained visual reasoning. Send ambiguous or high-stakes items to a larger multimodal model.
Low-cost reasoning tier in model cascades
With four effort levels and a vendor-reported 87.0 on AIME 2025, Mini can serve as the default tier, escalating to Lite or Pro only on failed validation or low confidence. This keeps most traffic on the lower-cost, faster tier. The cascade depends on a reliable confidence or validation signal, which you must build and test yourself.
Context-grounded long-document processing
The 262,144-token window lets Mini process long reports, transcripts or multi-document bundles in one request for summarization, field extraction and question answering over supplied text. Weak closed-book factuality makes provided context essential rather than optional. Check host billing above 128K input tokens, and validate recall on your own long documents.