Qwen3.8 27B

Mid-tier

Qwen3.8 27B is the download leader of the 2026 open-weight wave, pulling roughly 3.3 million times a month on Hugging Face, and the reason is the combination it hits. It is a dense 27 billion parameter native vision-language model taking text, image, and video input, released under plain Apache 2.0 with switchable thinking, and it fits on a single H100 at FP8 or a single RTX 4090 at 4-bit. The scores are unusual for the size: 84.3 on OSWorld-Verified, 90.3 on LiveCodeBench v6, 89.2 on GPQA Diamond, 73.0 on Terminal-Bench 2.1, and 61.7 on SWE-bench Pro. Native context is 262,144 tokens, extensible toward 1M. Hosted pricing runs $0.425 per million input tokens and $2.55 output. If you want one open model on your own hardware that does vision, agentic coding, and computer use without a multi-GPU bill, this is currently the default answer.

Input Price

$0.42

per 1M tokens

Output Price

$2.55

per 1M tokens

Context Window

tokens

Released

2026-08

Open source

Capabilities

textvisiontool-usecodereasoning

Key Strengths

✓Apache 2.0 with no field-of-use restrictions
✓Dense 27B, single-GPU at FP8 and consumer-GPU at 4-bit
✓Native vision-language: text, image, and video input
✓84.3 OSWorld-Verified and 90.3 LiveCodeBench v6 (vendor-reported)
✓Most-downloaded open model of the 2026 wave

Best For

▸Local and on-premise agents with vision
▸Computer-use automation on modest hardware
▸Fine-tuning base for domain-specific agents
▸Cost-sensitive high-volume inference

Pricing Details

Input tokens

$0.42

per 1M tokens

Output tokens

$2.55

per 1M tokens

Estimated cost per 1K requests

$1.70

~1K input + ~500 output tokens avg

Prices are subject to change. Check the official documentation for current pricing. See the cost calculator for detailed estimates.

Open Source Model

Qwen3.8 27B is free to download and self-host under the Apache-2.0. Hosted API pricing varies by provider (e.g., Together, Fireworks, Groq). See our open source LLM guide for deployment options.