Explore ProductsAI WorkspaceEnterpriseApps & SystemsManaged agentsWhite labelAPI
PlatformsDesktop AppChrome ExtensionWeb App
Explore ResourcesResourcesUse casesCase studies
Mission
SPIRITT logoSPIRITTQwenQwen3.8 27B

Qwen3.8 27B puts serious agent work in a practical size

A dense 27B open model now scores 52 on Artificial Analysis. SPIRITT serves it through U.S.-operated infrastructure, independently of Qwen's own cloud, with text, images, tools, 262K context, and $0.45/M input plus $3.20/M output reference pricing. Its real advantage is the combination of coding, vision, tools, and deployability, not a claim that it beats every larger frontier model.

Try Qwen3.8 27B Now
See how it works

Qwen3.8 27B across X

Open weights, day-zero serving, local coding-agent tests, independent evaluation, quantization work, and one very real overthinking caveat.

The practical member of the Qwen3.8 family

The Max release shows the ceiling. The dense 27B makes the generation usable on one capable machine and inexpensive hosted infrastructure.

Open weights, native vision, and a 262K route

Qwen3.8 27B is a dense 27B causal language model with a vision encoder, hybrid Gated DeltaNet and attention layers, Multi-Token Prediction, and Apache 2.0 weights. It understands text, images, and video in the native model card.

SPIRITT serves this Chinese-developed model through U.S.-operated infrastructure, independently of Qwen's own cloud, with text and image input, function calling, reasoning, a 262,144-token context, and reference pricing of $0.45/M input plus $3.20/M output.

Artificial Analysis scores the model at 52 with roughly 50 output tokens per second. It used about 160M output tokens in the Index run, so the compact parameter count does not automatically mean compact reasoning traces.

27B denseApache 2.0262K contextText + imageReasoning controls$0.45 / $3.20 per M

AA Intelligence Index v4.1.1

Claude Opus 5 max
63
Claude Fable 5 max
62
Grok 4.6 high
61
Qwen3.8-Max
58
GLM 5.3 Flash max
57
Qwen3.8 27B xhigh
52

Artificial Analysis model snapshots, August 29, 2026. A score of 52 is unusually strong for a dense 27B model, but it remains below the current overall leaders.

What a 27B model now gets done

Within-family gains are the cleanest vendor evidence. Independent quality, speed, and cost rows keep the practical tradeoffs visible.

Vendor terminal agents

Terminal-Bench 2.1

Claude Opus 4.6 Max
78.2%
Qwen3.8 27B
73%
Qwen3.7-Plus
64%
Qwen3.6 27B
63.4%
Muse Glimmer 30B
51.7%

Qwen model-card table. The 27B improves 9.6 points over its predecessor but still trails Opus 4.6 Max in the displayed cohort.

Vendor coding gain
Winner

DeepSWE v1.1

Qwen3.8 27B
42.2%
Qwen3.7-Plus
14.2%
Qwen3.6 27B
13.3%

Qwen-run benchmark. Winner means the displayed Qwen family cohort; no Opus result was reported in this table.

Vendor computer use
Winner

OSWorld-Verified

Qwen3.8 27B
84.3%
Qwen3.7-Plus
73.3%
Claude Opus 4.6 Max
72.7%
Qwen3.6 27B
63.9%

Qwen model-card table. The result is strong for this size, but OSWorld remains sensitive to agent setup.

Vendor instruction following
Winner

IFBench

Qwen3.8 27B
79.5%
Qwen3.7-Plus
79.1%
Muse Glimmer 30B
77%
Qwen3.6 27B
69.1%
Claude Opus 4.6 Max
62.5%

Qwen model-card table. The 27B narrowly leads Qwen3.7-Plus in the displayed cohort.

Vendor hard-knowledge limit

Humanity's Last Exam

Claude Opus 4.6 Max
40%
Qwen3.7-Plus
34.7%
Qwen3.8 27B
30.8%
Qwen3.6 27B
24%
Muse Glimmer 30B
22%

Qwen model-card table. The 27B improves over its predecessor and Muse Glimmer, while larger Qwen and Opus variants remain ahead.

Independent throughput

Output tokens / second

Grok 4.6 high
65.5
Qwen3.8 27B xhigh
50.4
GLM 5.3 Flash max
49.4
Qwen3.8-Max
21

Artificial Analysis model snapshots, August 29, 2026. Provider and hardware choices can move these rolling measurements.

Independent task economics

Cost per AA Intelligence task

GLM 5.3 Flash max
$0.09
Qwen3.8 27B xhigh
$0.37
Grok 4.6 high
$0.84
Qwen3.8-Max
$0.91

Artificial Analysis evaluation-route costs, August 29, 2026. Qwen 27B is inexpensive for its quality but not the cheapest displayed model.

Reference route price

Output price ($ / M tokens)

GLM 5.3 Flash
$0.50
Qwen3.8 27B
$3.20
Qwen3.8-Max
$6.00
Kimi K3
$15.00

Reference list prices for the exact hosted routes. Lower is better.

Independent source: Artificial Analysis Qwen3.8 27B model page and rolling model snapshots observed August 29, 2026. Vendor source: Qwen3.8-27B model card, including its harness footnotes and imported comparator rows. SPIRITT route facts were checked against current hosting documentation and the live model route. The model is remarkable for its size, not a replacement for repeated evaluation on your own high-stakes work.

How It Works

From a fresh workspace to a dense 27B model seeing the artifact, using tools, and checking its own work

01

Bring the real artifact

Open a SPIRITT workspace and add the codebase, screenshot, document, or research material the job actually depends on. The 262K route can hold substantial working context without a local model setup.

Open a SPIRITT workspace
02

Pick Qwen3.8 27B

Choose Qwen3.8 27B in the model picker. The SPIRITT route supports text, images, reasoning, and function calling, so it can inspect what it is building instead of treating vision as a separate tool.

Select Qwen3.8 27B in the model picker
03

Use medium effort before xhigh

Let it code, call tools, and verify outputs. Start routine work below xhigh so simple tasks do not burn excessive reasoning tokens, then raise effort only when the task earns it.

Build and automate with Qwen3.8 27B

Put a practical open model inside a full agent workspace

Qwen3.8 27B gets files, terminal, browser, image input, tools, memory, and a finish line without a local inference setup.

Try Qwen3.8 27B Now
Official model card
Questions

Qwen3.8 27B FAQ

01What is Qwen3.8 27B?+
Qwen3.8 27B is Qwen's August 2026 dense open-weight model with a vision encoder, 27B language-model parameters, hybrid attention, Multi-Token Prediction, and Apache 2.0 licensing.
02How is it different from Qwen3.8-Max?+
The 27B is the practical dense model: it can run on capable local hardware, and the SPIRITT route supports text and image input. Max is a much larger 2.4T MoE on SPIRITT, with higher quality and cost but slower independent throughput.
03Where does SPIRITT host Qwen3.8 27B?+
SPIRITT serves this Chinese-developed model through U.S.-operated infrastructure, independently of Qwen's own cloud. This describes the hosting operator and does not promise U.S.-only data residency.
04What is Qwen3.8 27B best for?+
Coding agents, screenshot and document workflows, computer-use tasks, local-model experimentation, and high-volume work where a dense open model is easier to operate than a trillion-parameter MoE.
05What are the main limitations of Qwen3.8 27B?+
It trails larger frontier models on hard terminal and knowledge tests, its xhigh default can overthink simple tasks, and Artificial Analysis measured high token use. The native context is 262K; 1M requires extension and additional memory.
06What are the reference token economics for Qwen3.8 27B?+
The hosted route's reference pricing is $0.45 per million input tokens and $3.20 per million output tokens. SPIRITT plan pricing and usage accounting are separate product terms.
07Can I connect a Qwen subscription from SPIRITT Subscriptions?+
Not currently. Choose Qwen3.8 27B through SPIRITT Models. The Subscriptions tab currently connects ChatGPT, Claude, and Grok accounts, not Qwen accounts.
08What is a good example of using Qwen3.8 27B well?+
Give it a frontend bug with the repository and screenshot: reproduce the issue, inspect the code, patch it, run tests, capture the rendered result, and compare the final screenshot against explicit acceptance criteria.
09Where is Qwen3.8 27B best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick Qwen3.8 27B in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt usingSPIRITT