Seed 1.8

Name: Seed 1.8
Author: ByteDance

Multimodal

ByteDance

Seed 1.8 is ByteDance's proprietary multimodal agent model, released 2026-02-17 and optimized for long-horizon agent scenarios with upgraded multimodal comprehension and more flexible context management. It accepts text, image, and video input and reasons before answering. The vendor report lists 94.3% on AIME 2025, 92.3% on MMLU, 84.9% on MMLU-Pro, 83.4% on MMMU, 81.3% on MathVista, 72.9% on SWE-Bench Verified, and 70.7% on AndroidWorld GUI control. DeepInfra serves it with a 256K-token context at $0.25 / $2.00 per 1M input/output tokens.

Key Specifications

Parameters

Context

256.0K

Release Date

February 17, 2026

Average Score

68.3%

API Documentation

Timeline

Key dates in the model's history

Announcement

February 17, 2026

Last Update

September 11, 2026

Today

September 19, 2026

Technical Specifications

Parameters

Training Tokens

Knowledge Cutoff

Family

Capabilities

MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)

$0.25

Output (per 1M tokens)

$2.00

Max Input Tokens

256.0K

Max Output Tokens

256.0K

Supported Features

Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

General Knowledge

Tests on general knowledge and understanding

MMLU

Official Table 1; think-high; no tools; Pass@1 • Self-reported

92.3%

Programming

Programming skills tests

SWE-Bench Verified

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

72.9%

Multimodal

Working with images and visual data

AI2D

Official Table 2; think-high; Pass@1 • Self-reported

89.1%

MathVista

Official Table 2; think-high; Pass@1 • Self-reported

87.7%

MMMU

Official Table 2; think-high; Pass@1 • Self-reported

83.4%

Other Tests

Specialized benchmarks

AetherCode

Official Table 1; think-high; no tools; Pass@1 • Self-reported

38.2%

AIME 2025

Official Table 1; think-high; no tools; Pass@1 • Self-reported

94.3%

AMO Bench

Official Table 1; think-high; no tools; Pass@1 • Self-reported

60.0%

AndroidWorld

Official Table 5; think-high; agentic GUI evaluation • Self-reported

70.7%

ARC-AGI

Official Table 1; think-high; no tools; Pass@1 • Self-reported

67.9%

Beyond AIME

Official Table 1; think-high; no tools; Pass@1 • Self-reported

77.0%

BFCL-V4

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

57.2%

BLINK

Official Table 2; think-high; Pass@1 • Self-reported

74.3%

BrowseComp

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

67.6%

BrowseComp-zh

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

81.3%

CharXiv-R

Official Table 2; think-high; Pass@1 • Self-reported

71.4%

CountBench

Official Table 2; think-high; Pass@1 • Self-reported

96.3%

DUDE

Official Table 2; think-high; Pass@1 • Self-reported

69.4%

DynaMath

Official Table 2; think-high; Pass@1 • Self-reported

61.5%

EMMA

Official Table 2; think-high; Pass@1 • Self-reported

60.9%

ERQA

Official Table 2; think-high; Pass@1 • Self-reported

58.8%

Hallusion Bench

Official Table 2; think-high; Pass@1 • Self-reported

63.9%

Humanity's Last Exam (with tools, text-only)

Official Table 4; think-high; agentic search; text-only with tools • Self-reported

40.9%

LiveCodeBench Pro

Official Table 1; think-high; no tools; Elo rating • Self-reported

64.3%

LiveCodeBench v6

Official Table 1; think-high; no tools; Pass@1 • Self-reported

79.5%

LiveSports-3K

Official Table 3; think-high; public video evaluation • Self-reported

77.5%

LongVideoBench

Official Table 3; think-high; public video evaluation • Self-reported

77.4%

LVBench

Official Table 3; think-high; public video evaluation • Self-reported

73.0%

MathVision

Official Table 2; think-high; Pass@1 • Self-reported

81.3%

Minerva

Official Table 3; think-high; public video evaluation • Self-reported

62.4%

MMLU-Pro

Official Table 1; think-high; no tools; Pass@1 • Self-reported

84.9%

MMMU-Pro

Official Table 2; think-high; Pass@1 • Self-reported

73.2%

MMStar

Official Table 2; think-high; Pass@1 • Self-reported

79.9%

MMVU

Official Table 3; think-high; public video evaluation • Self-reported

73.1%

MotionBench

Official Table 3; think-high; public video evaluation • Self-reported

70.6%

MuirBench

Official Table 2; think-high; Pass@1 • Self-reported

78.7%

Multi-Challenge

Official Table 1; think-high; no tools; Pass@1 • Self-reported

66.7%

Multi-SWE-Bench

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

42.0%

OmniDocBench 1.5

Official Table 2; think-high; normalized edit distance (lower is better) • Self-reported

10.6%

OSWorld

Official Table 5; think-high; agentic GUI evaluation • Self-reported

61.9%

OVBench

Official Table 3; think-high; public video evaluation • Self-reported

65.1%

OVOBench

Official Table 3; think-high; public video evaluation • Self-reported

72.6%

PHYBench

Official Table 1; think-high; no tools; Pass@1 • Self-reported

41.0%

RefSpatialBench

Official Table 2; think-high; Pass@1 • Self-reported

56.3%

SimpleVQA

Official Table 2; think-high; Pass@1 • Self-reported

65.4%

SuperGPQA

Official Table 1; think-high; no tools; Pass@1 • Self-reported

64.8%

TempCompass

Official Table 3; think-high; public video evaluation • Self-reported

86.9%

Terminal-Bench 2.0

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

45.2%

TOMATO

Official Table 3; think-high; public video evaluation • Self-reported

60.8%

TVBench

Official Table 3; think-high; public video evaluation • Self-reported

71.5%

VideoMME w sub.

Official Table 3; think-high; subtitles included • Self-reported

87.8%

VideoMMMU

Official Table 3; think-high; public video evaluation • Self-reported

82.7%

VideoSimpleQA

Official Table 3; think-high; public video evaluation • Self-reported

67.8%

VLMsAreBiased

Official Table 2; think-high; Pass@1 • Self-reported

62.0%

VLMsAreBlind

Official Table 2; think-high; Pass@1 • Self-reported

93.0%

WideSearch

Official Table 4; think-high; benchmark-specific agent tools • Self-reported

63.8%

ZEROBench

Official Table 2; think-high; Pass@1 • Self-reported

11.0%

License & Metadata

License

proprietary

Announcement Date

February 17, 2026

Last Updated

September 11, 2026

Similar Models

All Models

Seed 2.0 Pro

ByteDance

Best score:0.9 (GPQA)

Released:Feb 2026

Price:$0.50/1M tokens

Seed 2.0 Lite

ByteDance

Best score:0.9 (GPQA)

Released:Feb 2026

Seed 2.1 Pro

ByteDance

Released:Jun 2026

Seed 2.1 Turbo

ByteDance

Released:Jun 2026

Seed 2.0 Mini

ByteDance

Best score:0.8 (GPQA)

Released:Feb 2026

Price:$0.10/1M tokens

Seedance 1.5 Pro

ByteDance

Released:Jan 1970

ERNIE 5.0

Baidu

Best score:0.8 (GPQA)

Released:Jan 2025

Kimi-k1.5

Moonshot AI

Best score:0.9 (MMLU)

Released:Jan 2025

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.