Text and chat · Alibaba (Qwen)

Qwen open models

Qwen models with open weights, run in private mode. Qwen 3.8 27B reads images and video at a low price.

Private Moderate cost
Latest versionQwen3.8 27BAugust 14, 2026
Artificial Analysis Intelligence Index40 points#23 of 90 on Artificial Analysis Intelligence Index
Price$1.439 per million tokens written$0.00101 a page written
Key spec262K tokens of context

What sets it apart

Qwen’s open models cover many sizes, from phone-class 7B to a 2.4T flagship, mostly under Apache 2.0. The 3.8 generation brings a 27B model scoring 34 on the Artificial Analysis index.

  • Sizes from 7B to 2.4T
  • Apache 2.0 on most models
  • 27B model with image and video input

Recommended for: get quick answers for very little.

Strengths and limits

Where it shines

  • Private mode on every version
  • Wide choice of sizes, from very cheap to very large
  • Qwen 3.8 27B reads images and video, 262K context

Think twice when

  • Older 2507, Next and Coder models trail the 3.6 to 3.8 ones
  • Qwen 2.5 7B (encrypted) has no tool calling, so none of Sub Rosa's agent tools
  • Qwen 3.8 27B: 262K context, vision, tool calling
  • Open weights

Use cases

Running a model on your own machine

Explain in plain words the difference between a mortgage with a fixed rate and one with a variable rate, with a numeric example.

Qwen3.8 27B has open Apache 2.0 weights and is small enough to self-host.

Reading a diagram or a form

This is a scanned tax form. List each field and what is filled in, and mark the empty ones.

Qwen3 VL and the 3.6 and 3.8 models accept images.

Code generation

Write a SQL query that returns the ten customers with the highest total orders in 2025, and explain it line by line.

The Coder 480B model is tuned for programming tasks.

Against its rivals

  • GemmaGemma 4 31B scores 15 on the index, about half of Qwen3.8 27B at 34.#71 of 90 on Artificial Analysis Intelligence Index (17 points)
  • GPT OSSgpt-oss-120b scores 12 and is older, but runs fast on one GPU.#79 of 90 on Artificial Analysis Intelligence Index (12 points)
  • DeepSeekDeepSeek V4.1 Flash scores 39 with a 1M window but is a much larger download.#24 of 90 on Artificial Analysis Intelligence Index (39 points)
Compare side by side ↗

Version history

Each version, dated, with what it changed over the one before.

Artificial Analysis Intelligence Index, version after versionHigher is better · Artificial Analysis

Hollow points are earlier versions no longer offered in Sub Rosa.

See the numbers
VersionReleasedpoints
Qwen3 2507 and Coder2025-0713
Qwen3 Next and Qwen3 VL2025-0913
Qwen3.5 small models2026-0219
Qwen3.5 9B2026-0311
Qwen3.62026-0421
Qwen3.8 27B2026-08-1434
Qwen3.8 2.4T A95B2026-0840
  1. Qwen2.5 7B
    • Dense 7B model of the Qwen2.5 family, Apache 2.0, up to 128K tokens of input and 8K of output.
    Source : qwenlm.github.io
  2. Qwen3 2507 and Coder
    • 235B mixture of experts (22B active) in instruct and thinking versions with a 256K window, plus a 480B coder model (35B active).
    Source : artificialanalysis.ai
  3. Qwen3 Next and Qwen3 VL
    • Next: 80B mixture of experts with only 3B active. VL: vision-language models that read images, 235B and 30B sizes.
    Source : artificialanalysis.ai
  4. Qwen3.5 small models
    • 35B mixture of experts with 3B active, 262K context, image input, Apache 2.0.
    Source : artificialanalysis.ai
  5. Qwen3.5 9B
    • Dense 9.7B model reading text, images and video, with a 262K window.
    Source : artificialanalysis.ai
  6. Qwen3.6
    • 27.8B dense and 36B mixture of experts (3B active) models, Apache 2.0, text, image and video input.
    Source : artificialanalysis.ai
  7. Qwen3.8 2.4T A95B
    • Open-weights flagship: 2.4T parameters with 95B active, scoring 40 on the index, under the Qwen3.8-Max license.
    Source : artificialanalysis.ai
  8. Qwen3.8 27BLatest
    • GPQA Diamond rises from 87.8 to 89.2 and SWE-bench Pro from 53.5 to 61.7 over Qwen3.6 27B, per the model card.
    • Adjustable thinking depth and stronger performance on long agent tasks.
    Source : huggingface.co

Benchmarks

How to read these scores: what a benchmark tells you.

Artificial Analysis Intelligence Index: #23 of 90 models measured in the catalog. One score averaging about ten independent tests of reasoning, knowledge, math, coding and agent work, all run the same way for every model. Higher is better. Each model is listed at its highest reasoning effort. Two or three points apart is a close call; ten points is a different class of model.

Artificial Analysis Intelligence IndexHigher is better · Artificial Analysis, October 4, 2026
  1. 1. Claude Opus 5.5Claude58
  2. 2. Claude Sonnet 5.5Claude56
  3. 3. Claude Fable 5.1Claude53
  4. 4. GPT-6 AstraGPT53
  5. 5. Gemini 4 ArgonGemini53
  6. 6. GPT-6.1 SolGPT52
  7. 7. Muse WonderMuse52
  8. 8. Claude Opus 5Claude51
  9. 9. Claude Fable 5Claude50
  10. 10. GPT-6 SolGPT48
  11. 23. Qwen 3.8 2.4TQwen open models40
  12. 38. Qwen 3.8 27BQwen open models34
  13. 66. Qwen 3.6 27BQwen open models21
See the numbers
Modelpoints
1. Claude Opus 5.558
2. Claude Sonnet 5.556
3. Claude Fable 5.153
4. GPT-6 Astra53
5. Gemini 4 Argon53
6. GPT-6.1 Sol52
7. Muse Wonder52
8. Claude Opus 551
9. Claude Fable 550
10. GPT-6 Sol48
23. Qwen 3.8 2.4T40
38. Qwen 3.8 27B34
66. Qwen 3.6 27B21

Announced by the maker

Measured by the maker with its own settings: useful, but not directly comparable with the independent scores above.

BenchmarkModelScoreSource
gpqa-diamondQwen 3.8 27B89.2huggingface.co

All 15 models in this family

ModelUsePrivacyPrice
Qwen 3.8 27Bqwen-3-8-27bTextPrivate$1.439 per million tokens written$0.00101 a page written
Qwen 3.8 2.4Tqwen-3-8-2-4t-a95bTextPrivate$3.372 per million tokens written$0.00236 a page written
Qwen 3.6 27Bqwen3-6-27bTextPrivate$1.461 per million tokens written$0.00102 a page written
Qwen 3.6 35B A3Bqwen3-6-35b-a3bTextPrivate$0.45 per million tokens written$0.000315 a page written
Qwen 3.5 35B A3Bqwen3-5-35b-a3bTextPrivate$0.45 per million tokens written$0.000315 a page written
Qwen 3.5 9Bqwen3-5-9bTextPrivate$0.0674 per million tokens written$0.0000472 a page written
Qwen 3 235B A22B Instruct 2507qwen3-235b-a22b-instruct-2507TextPrivate$0.337 per million tokens written$0.000236 a page written
Qwen 3 235B A22B Thinking 2507qwen3-235b-a22b-thinking-2507TextPrivate$1.574 per million tokens written$0.0011 a page written
Qwen 3 Coder 480B Turboqwen3-coder-480b-a35b-instruct-turboTextPrivate$0.674 per million tokens written$0.000472 a page written
Qwen 3 Next 80bqwen3-next-80bTextPrivate$0.854 per million tokens written$0.000598 a page written
Qwen3 VL 235Bqwen3-vl-235b-a22bTextPrivate$0.854 per million tokens written$0.000598 a page written
Qwen 3.8 27B E2EEe2ee-qwen3-8-27bTextPrivate$1.587 per million tokens written$0.00111 a page written
Qwen 3.6 35B A3B FP8 E2EEe2ee-qwen3-6-35b-a3bTextPrivate$0.531 per million tokens written$0.000371 a page written
Qwen3 VL 30B A3B E2EEe2ee-qwen3-vl-30b-a3b-pTextPrivate$0.405 per million tokens written$0.000283 a page written
Qwen 2.5 7B E2EEe2ee-qwen-2-5-7b-pTextPrivate$0.0584 per million tokens written$0.0000409 a page written

Specifications

SpecificationQwen3.8 2.4T A95BQwen3.8 27BQwen3.6 35B A3BQwen3 235B A22B 2507
ReleasedAugust 2026August 14, 2026April 2026July 2025
Parameters2.4T total, 95B active27B35B total, 3B active235B total, 22B active
WeightsOpen, Qwen3.8-Max License (commercial use with conditions)Open, Apache 2.0Open, Apache 2.0Open, Apache 2.0
Acceptstexttext, images, videotext, images, videotext
Producestexttexttexttext
Context262K tokens262K tokens256K tokens128K tokens
AbilitiesUses Sub Rosa’s tools, Thinks before answering, Tuned for codeReads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for codeReads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for codeUses Sub Rosa’s tools
Price$3.372 per million tokens written$1.439 per million tokens written$0.45 per million tokens written$0.337 per million tokens written
PrivacyPrivatePrivatePrivatePrivate

Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.