Qwen3.8 27B

Aug 2026
Qwen3.8 27B
Compare in
on
$1.01per task?
$0.42 / $2.55per 1M in / outcache $0.085
114 tok/sfirst token in 1.2 s?
262Kcontext

Rankingat xhigh effort?

Full ranking →
1Sonnet 5.563.62Opus 5.559.62GPT-6 Astra59.64Gemini 4 Argon57.14Muse Wonder Full57.1
···
43Muse Spark 1.16.144Qwen 27B5.645Gemini 3.1 Pro Preview4

Strong at

S
Science#30 of 312Long context#27 of 273
A
Agents & work tasks#38 of 175Reasoning#66 of 334
B
Coding#44 of 132Knowledge#92 of 318Vision#41 of 124

Alternatives?

ModelScoreCostPer taskWhat changes
CheaperGPT-6.1 Solmax×10.1-28%$0.72Remove unsupported parameters+5BetterClaude Haiku 5.5low×2.3-98%$0.024Remove sampling parameters+5IOpenK2 Horizon 7Bthinking-82%-45%$0.30 / $1.20No image inputCheaperClaude Sonnet 5.5high×7.9-12%$0.88No forced tool use+5CheaperGPT-6 Astralow×7.5-19%$0.82Monitoring can return HTTP 403+2

Run it yourself

estimate
Fits on✗ 8 GB (phone, laptop)✗ 16 GB✓ 24 GB GPU · compressed✓ 48 GB GPU · 8-bit✓ 80 GB GPU · full quality
LicenceApache-2.0free for commercial use

History

3 versions
0246Jul 26Release dateTerminal-Bench 4.03.6 27B3.8 27B
0246Jul 26Release dateTerminal-Bench 4.03.6 27B3.8 27B

Benchmarks

in Coding
Terminal-Bench 2.179.8#23
Terminal-Bench 4.05.6#44
SciCode46.6#67
FrontierCode 1.134.2#22
Terminal-Bench Science 0.10.5#25
SWE-bench Pro (edition unstated)lab61.7
DeepSWE 1.1lab42.2

Watch out 5

max_tokens range · QwenCloud›

With thinking enabled, max_tokens outside 1–32768 returns HTTP 400; use max_completion_tokens to budget reasoning plus answer without that cap. Source ↗

Static YaRN affects shorter texts · Self-hosted›

Static YaRN can hurt short-text performance; change rope_parameters only when longer contexts are required. Source ↗

tool_choice only supports auto or none · QwenCloud›

With thinking enabled, tool_choice supports only auto or none. Source ↗

Include the word JSON · QwenCloud›

JSON Object mode requires the word JSON in a system or user message; otherwise the API returns an error. Source ↗

Agent search_strategy not supported · QwenCloud / Chat Completions›

Chat Completions does not support Qwen3.8 agent search_strategy; use the Responses web_search tool for multi-turn retrieval. Source ↗

FAQ

How much does Qwen3.8 27B cost?›

$0.42 per 1M input tokens and $2.55 per 1M output tokens; cached input $0.085 per 1M; cheapest via Tenstorrent at $0.02 / $0.50. A typical benchmark task costs about $1.01 at xhigh effort.

Is Qwen3.8 27B free to use?›

There is no free API; the cheapest paid price is $0.02 per 1M input tokens. The weights are open, so you can run it yourself.

How good is Qwen3.8 27B, and what is it best at?›

#66 of 334 in Chat & Knowledge, #44 of 132 in Coding, #38 of 133 in Agents & Tool Use, #91 of 195 in Writing & Language, #41 of 124 in Vision & Documents. Strongest at: Agents & work tasks, Reasoning, Science, Long context.

Is Qwen3.8 27B better than Qwen3.6 27B?›

On Terminal-Bench 4.0, Qwen3.8 27B scores 5.6 vs 0 — higher. It is 62% more expensive per task.

How can I use Qwen3.8 27B?›

Through the API of AkashML, Alibaba, Cerebras, Chutes, Cloudflare and 16 more.

What is the context window of Qwen3.8 27B?›

262K tokens, with up to 131K output tokens.

When was Qwen3.8 27B released? Is there a newer version?›

Released 2026-08-14. It is the current version in its line.

Is Qwen3.8 27B open source?›

Yes, the weights are open (Apache-2.0 licence).

Can I run Qwen3.8 27B locally?›

The weights are open and can be self-hosted; hardware needs are not confirmed yet.

Can Qwen3.8 27B read images?›

It accepts text and images, video as input.

What is Qwen3.8 27B?›

Qwen3.8 27B is an open-weights language model from Alibaba (Qwen), released Aug 2026, the successor to Qwen3.6 27B.

What are the best alternatives to Qwen3.8 27B?›

GPT-6.1 Sol (cheaper), Claude Haiku 5.5 (better), K2 Horizon 7B (open weights).