Qwen3.8-27B

How it is built

Qwen3.8-27B is the dense model of the family. All 27B parameters work on every token.¹¹1Qwen, Qwen3.8-27B model card, August 2026. SWE-bench Pro and DeepSWE 1.1 with the Claude Code harness, temperature 1.0, top_p 0.95 and a 256K context window; the Terminal Bench 2.1 row is labelled "Terminus"; video understanding "from STEM diagrams and documents to hour-scale videos". https://huggingface.co/Qwen/Qwen3.8-27B Qwen3.8-Max and Qwen3.8-Flash, the two larger open models, send each token through a handful of experts instead.²²2Qwen, Qwen3.8-2.4T-A95B model card: 2.4T parameters with 95B active, layout 23 x (3 x Gated DeltaNet, 1 x Gated Attention), thinking that cannot be disabled, Qwen3.8-Max License. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B ³³3Qwen, Qwen3.8-Flash-Next model card: 125B parameters with 6B active, Qwen Community License 1.0. https://huggingface.co/Qwen/Qwen3.8-Flash-Next

The layer pattern is theirs. The 64 layers form 16 blocks of three Gated DeltaNet layers and one Gated Attention layer.¹¹²²³³ The hidden size is 5,120, and each feed-forward block is 17,408 wide. Attention uses 24 query heads and 4 key-value heads.¹¹ The padded vocabulary holds 248,320 tokens, and training added multi-token prediction.¹¹

A vision encoder of 27 layers handles images and video.⁴⁴4Qwen, Qwen3.8-27B config.json (text dtype bfloat16, vision encoder depth 27). Arithmetic: 27B x 2 bytes = 54 GB, before the vision encoder and the KV cache. https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json Qwen calls the result a native vision-language model and says it follows videos an hour long.¹¹

What is new

Qwen released it on 14 August 2026, two days after the weights of the 2.4T model.⁵⁵5Qwen, Qwen3.8 repository README, news entries of 12 August 2026 (Qwen3.8-2.4T-A95B weights) and 14 August 2026 (Qwen3.8-27B). https://github.com/QwenLM/Qwen3.8

Against Qwen3.6-27B, the card reports 61.7 on SWE-bench Pro where the older model had 53.5, both in the Claude Code harness at temperature 1.0 with a 256K context.¹¹ With the Terminus harness, Terminal Bench 2.1 went from 63.4 to 73.0.¹¹ NL2Repo-Bench asks for a whole repository from a written description. There the score is 42.3, against 36.2.¹¹ On DeepSWE 1.1 the older model resolved 13.3 and the new one 42.2.¹¹

Thinking is on by default and can be turned off for a single request.¹¹ The open 2.4T model has no such switch.²² Depth is tuned with reasoning_effort, and preserve_thinking keeps the reasoning of earlier turns in context.¹¹

Running it

This is the one open Qwen3.8 model under Apache 2.0.¹¹ The two larger ones carry Qwen's own licences, with conditions on commercial use.²²³³

In BF16, at two bytes per parameter, the language model's weights take about 54 GB.⁴⁴ The FP8 checkpoint halves that to about 27 GB.⁶⁶6Qwen, Qwen3.8-27B-FP8 model page. Arithmetic: 27B x 1 byte = 27 GB. https://huggingface.co/Qwen/Qwen3.8-27B-FP8

The context is 262,144 tokens natively and 1,000,000 with YaRN scaling.¹¹ Qwen has announced a hosted version with a 1M context by default. The card still lists it as coming soon.¹¹