FrankX.AI
Alibaba (Qwen)GA

Qwen3.8-27B

Apache 2.0 multimodal weights small enough to run yourself — the licence is the story.

Context

262K

Max output

—

Input /1M

Open

Output /1M

Open

Best for

  • Self-hosted and air-gapped deployments
  • Commercial use without a revenue-share clause
  • Local multimodal work (text, image, video in) at 27.8B

Watch out

Alibaba’s published comparison leads on all four benchmarks it chose to report — the shape of a vendor-selected set. No first-party hosted API pricing exists yet, so “free” means self-hosted, and you pay in GPUs.

For creators. The private lane: run creator workloads on your own hardware when the content should never leave the building.

Benchmarks

terminal bench 2 173
swe bench pro61.7
gpqa diamond89.2
ifbench79.5

Capabilities

  • Apache 2.0 open weights - commercial use with no revenue-share clause
  • 27.8B dense multimodal: text, image and video input
  • 262,144-token native context (Alibaba states extension to 1M)
  • Runs locally - the capability-per-GB story, not the frontier story

More from Alibaba (Qwen)

Qwen3.7-Max

Sources