vs
4:9
overall capability wins (1 ties across 14 scored). Qwen3.8 27B leads overall.
- Reasoning2:4
- Coding0:3
- Arena1:0
- Tool use0:1
- Composite indices0:2
- Specs1:1
Verdict
Qwen3.8 27B leads coding (widest gap: +194.4 on WebDev Elo); Qwen3.5 397B-A17B ranks higher on LMArena (+5 Elo); Qwen3.8 27B allows ~4.0x more max output (131K vs 33K).
- Qwen3.8 27B leads coding (widest gap: +194.4 on WebDev Elo)
- Qwen3.5 397B-A17B ranks higher on LMArena (+5 Elo)
- Qwen3.8 27B allows ~4.0x more max output (131K vs 33K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Qwen3.5 397B-A17B
- $0.70
- Qwen3.8 27B
- -
Price per 1M tokens
Full comparison
Organization
Alibaba
Alibaba
Family
Qwen
Qwen
License
Apache 2.0
Apache 2.0
Open weights
Yes
Yes
Release
Mar 1, 2026
Aug 14, 2026
Knowledge cutoff
-
-
API / provider
Together / Fireworks (ref.)
No primary API price listed
Modalities
text → text
text, image, video → text
Specs
Context window
262K
262K
Max output
33K
131K
Parameters
397B (17B act.)
27B
Pricing
Input $/1M
$0.40
—
Output $/1M
$1.20
—
Blended $/1M (3∶1)
$0.60
-
Speed
tok/s
75
-
TTFT (s)
0.4
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
89.3
90.5
Humanity's Last Exam
29.0
33.9
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
30.8
15.6
AA-LCR v1.1
77.3
82.0
CritPt
1.7
5.4
MMMU-Pro
77.3
76.3
IFBench
78.8
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
61.7
SWE-bench Multilingual
—
—
LiveCodeBench
—
90.3
Terminal-Bench 2.1
51.3
79.8
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
44.8
46.6
CursorBench
—
—
SWE-Rebench
59.9
—
NL2Repo-Bench
—
42.3
DeepSWE
—
42.2
WebDev Arena
1,399
1,593
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,442
1,437
| Metric | Qwen3.5 397B-A17B | Qwen3.8 27B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Alibaba | Alibaba | — |
| Family | Qwen | Qwen | — |
| License | Apache 2.0 | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Mar 1, 2026 | Aug 14, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Together / Fireworks (ref.) | No primary API price listed | — |
| Modalities | text → text | text, image, video → text | — |
| Specs | |||
| Context window | 262K | 262K | tie |
| Max output | 33K | 131K | B +98K |
| Parameters | 397B (17B act.) | 27B | A +370 |
| Pricing | |||
| Input $/1M | $0.40 | — | — |
| Output $/1M | $1.20 | — | — |
| Blended $/1M (3∶1) | $0.60 | - | — |
| Speed | |||
| tok/s | 75 | - | — |
| TTFT (s) | 0.4 | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 89.3 | 90.5 | B +1.2 pts |
| Humanity's Last Exam | 29.0 | 33.9 | B +4.9 pts |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 30.8 | 15.6 | A +15.2 pts |
| AA-LCR v1.1 | 77.3 | 82.0 | B +4.7 pts |
| CritPt | 1.7 | 5.4 | B +3.7 pts |
| MMMU-Pro | 77.3 | 76.3 | A +1.0 pts |
| IFBench | 78.8 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | 61.7 | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 90.3 | — |
| Terminal-Bench 2.1 | 51.3 | 79.8 | B +28.5 pts |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 44.8 | 46.6 | B +1.8 pts |
| CursorBench | — | — | — |
| SWE-Rebench | 59.9 | — | — |
| NL2Repo-Bench | — | 42.3 | — |
| DeepSWE | — | 42.2 | — |
| WebDev Arena | 1,399 | 1,593 | B +194 Elo |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,442 | 1,437 | A +5 Elo |
Qwen3.5 397B-A17B
- Context
- 262K / 33K out
- Parameters
- 397B (17B act.)
- Price
- $0.40 / $1.20
- Speed
- 75 tok/s · 0.4s TTFT
- Modalities
- text → text
- License
- Apache 2.0
Large open MoE from the Qwen3.5 line under Apache 2.0.
Qwen3.8 27B
- Context
- 262K / 131K out
- Parameters
- 27B
- Price
- —
- Speed
- —
- Modalities
- text, image, video → text
- License
- Apache 2.0
Open-weight Qwen3.8 27B dense vision-language model (Apache-2.0, released Aug 14, 2026) — native image and video understanding, 262K context extensible to 1M, flexible thinking control via reasoning_effort, and strong agentic coding for its size (Terminal-Bench 2.1 73.0, SWE-bench Pro 61.7, GPQA Diamond 89.2).
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.