# Seed 1.6

Seed 1.6 combines sparse MoE efficiency with adaptive reasoning in one model. It switches between full extended thinking, direct response, and automatic mode selection.

- **Model ID:** `bytedance/seed-1.6`
- **Type:** chat
- **Providers:** bytedance
- **Context window:** 256,000
- **Maximum output tokens:** 32,000
- **Pricing:** $0.25/1M input tokens, $2/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/seed-1.6

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'bytedance/seed-1.6',
  prompt: 'Why is the sky blue?'
})
```

## About

Seed 1.6 uses a sparse Mixture-of-Experts (MoE) architecture with 23B active parameters drawn from 230B total. It builds on MoE design principles from ByteDance's Seed 1.5 research. Only the parameter subset most relevant to each token activates during inference, which keeps costs closer to a 23B dense model than a 230B one.

Seed 1.6 offers three reasoning modes. `FullCoT` enables extended chain-of-thought for demanding analytical problems. `NoCoT` produces direct completions without a visible reasoning trace, suited to latency-sensitive applications. `AdaCoT`, the adaptive mode, decides whether reasoning tokens are warranted based on inferred question difficulty. It saves compute on simple queries while engaging deeper reflection when the problem demands it. A parallel decoding enhancement also generates additional thinking tokens without retraining, yielding eight-point improvements on the BeyondAIME benchmark.

The context window of 256K tokens supports long-document processing and multimodal input. On China's 2025 Gaokao, Seed 1.6 scored 683/750 in humanities (ranked first) and 648/750 in science (ranked second), rising to 676/750 with higher-resolution images. On India's JEE Advanced entrance exam, the model placed in the top 10 with 100% accuracy in math across all five sampling rounds.

See https://console.byteplus.com/ark/region:ark+ap-southeast-1/model/detail?Id=seed-1-6 for the full write-up, tables, and methodology.

## What to consider

If your workload relies on extended thinking, confirm that your chosen provider exposes the full reasoning token budget without truncation. Compare token pricing ($0.25 in, $2 out per million tokens when listed).

## When to use

### Best For

- Long-document analysis where the context window of 256K tokens lets you ingest entire contracts, research papers, or codebases in one prompt
- Mixed-difficulty workloads where `AdaCoT` calibrates compute spend automatically, routing complex queries to full reasoning and simple ones to direct response
- Multimodal pipelines that combine text and visual data in the same request through integrated Vision-Language Model (VLM) capability
- GUI-based interaction tasks that require understanding screenshots or interface layouts alongside natural language instructions
- Competitive mathematics, science, and humanities domains where benchmark results confirm consistent performance

### Consider Alternatives When

- Your task is purely text generation with no visual input, and a dense model may offer more predictable routing behavior
- You need more than 32K tokens in a single response
- Your pipeline demands real-time streaming with deterministic latency and can't tolerate variable thinking token overhead from `FullCoT` or `AdaCoT`
- Cost per token is the primary constraint and a smaller distilled model meets your quality bar

## Best for

- **Long-document analysis:** The context window of 256K tokens lets you ingest entire contracts, research papers, or codebases in one prompt
- **Mixed-difficulty workloads:** `AdaCoT` calibrates compute spend automatically, routing complex queries to full reasoning and simple ones to direct response
- **Multimodal pipelines:** Integrated VLM capability combines text and visual data in the same request
- **GUI-based interaction:** Tasks that require understanding screenshots or interface layouts alongside natural language instructions
- **Competitive academic domains:** Mathematics, science, and humanities workloads where benchmark results confirm consistent performance

## Consider alternatives

- **Pure text generation:** A dense model may offer more predictable routing behavior when no visual input is involved
- **Output length limits:** Requests needing more than 32K tokens in a single response exceed capacity
- **Deterministic latency:** Pipelines can't tolerate variable thinking token overhead from `FullCoT` or `AdaCoT`
- **Cost-sensitive workloads:** A smaller distilled model may meet your quality bar at lower cost per token

## Frequently asked questions

### What are the three reasoning modes in Seed 1.6 and when should I use each?

The three modes are `FullCoT`, `NoCoT`, and `AdaCoT`. `FullCoT` generates an extended chain-of-thought trace before answering, best for complex multi-step problems. `NoCoT` skips the reasoning trace for direct, low-latency responses. `AdaCoT` selects between the two based on estimated question difficulty, making it the practical default for mixed workloads.

### Does Seed 1.6 support image inputs?

Yes. Seed 1.6 incorporates Vision-Language Model (VLM) capabilities. You can process both text and visual data in the same request.

### What is parallel decoding in the context of Seed 1.6?

It's a training-free inference enhancement that generates additional thinking tokens without changing the base model weights. It deepens reasoning capacity at inference time, yielding eight-point improvements on the BeyondAIME benchmark.

### How does the sparse MoE architecture affect cost?

Only 23B of the 230B total parameters activate per forward pass. Inference compute stays closer to a 23B dense model than a 230B one. This lowers per-token cost while retaining the representational capacity of the full parameter count.

### What academic benchmarks were used to evaluate Seed 1.6?

ByteDance evaluated Seed 1.6 on China's 2025 Gaokao (683/750 in humanities, ranked first; 648/750 in science, ranked second, rising to 676/750 with higher-resolution images) and India's JEE Advanced entrance exam (top-10 placement, 100% math accuracy across five sampling rounds). See https://console.byteplus.com/ark/region:ark+ap-southeast-1/model/detail?Id=seed-1-6 for the full tables.

### Is Seed 1.6 available for commercial use through AI Gateway?

Yes. You can access Seed 1.6 through ByteDance via AI Gateway with an API key or OIDC token. You don't manage upstream provider credentials yourself.

## Links

- [Model page](https://vercel.com/ai-gateway/models/seed-1.6)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://console.byteplus.com/ark/region:ark+ap-southeast-1/model/detail?Id=seed-1-6)
- [Provider pricing](https://docs.byteplus.com/en/docs/ModelArk/1544106)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.