Coding Horizon

Qwen 3.8 27B Just Made GPT 5.6 Look Wasteful Now

Every figure this video puts on screen, with where it comes from.

The model

Claim Value Source
Parameters 27B Qwen/Qwen3.8-27B model card
Licence Apache 2.0 Qwen/Qwen3.8-27B model card
Native context 262,144 tokens Qwen/Qwen3.8-27B model card
Extended context up to 1,000,000 tokens, via YaRN Qwen/Qwen3.8-27B model card
Modality native vision language model, images and video Qwen/Qwen3.8-27B model card
Thinking mode on by default, can be disabled per request Qwen/Qwen3.8-27B model card
Reasoning effort reasoning_effort, with low, medium and xhigh Qwen/Qwen3.8-27B model card
Weights on disk 55.6 GB of BF16 safetensors across 18 shards files and versions
Published 14 August 2026, by the Qwen team at Alibaba Cloud Qwen/Qwen3.8-27B model card

The model card’s own framing of what it is for: “comprehensive improvements across coding, professional work, research, and long-horizon agentic tasks”, and “stronger autonomous planning and better handling of environment feedback, leading to more reliable end-to-end task completion”.

The head to head

Artificial Analysis, comparing Qwen3.8 27B at xhigh against GPT 5.6 Luna at xhigh:

Metric Qwen3.8 27B (xhigh) GPT 5.6 Luna (xhigh) Better
Intelligence Index 52 50 Qwen
Blended price, per 1M tokens $0.43 $0.17 GPT
Output speed, tokens per second 54 145 GPT
Time to first token 2.98s 52.74s Qwen
Context window 256k 1000k GPT

Source: Artificial Analysis, Qwen3.8 27B vs GPT-5.6 Luna (xhigh).

The score depends on which GPT 5.6 Luna configuration is compared. At low reasoning effort the same index puts Luna at 34, against Qwen3.8 27B’s 52: comparison. The video compares the xhigh configuration in both directions, which is the closest of the published matchups.

The Hugging Face ranking

Qwen3.8 27B reached number one on Hugging Face’s global trending models board on 17 August 2026, three days after release, and passed three million downloads in the same window.

What is illustrative rather than measured

Some shots draw a shape rather than a published figure, and are labelled on screen as what they are rather than carrying a number:

Not verified