vs
1:5
overall capability wins (0 ties across 6 scored). Qwen3.8 27B leads overall.
- Reasoning0:2
- Arena0:1
- Specs1:2
Verdict
Qwen3.8 27B edges reasoning & knowledge; Qwen3.8 27B ranks higher on LMArena (+176 Elo); Qwen3.8 27B offers ~8.0x larger context (262K vs 33K).
- Qwen3.8 27B edges reasoning & knowledge
- Qwen3.8 27B ranks higher on LMArena (+176 Elo)
- Qwen3.8 27B offers ~8.0x larger context (262K vs 33K)
- Qwen3.8 27B allows ~16.0x more max output (131K vs 8K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Qwen2 72B Instruct
- $0.45
- Qwen3.8 27B
- -
Price per 1M tokens
Full comparison
Organization
Alibaba
Alibaba
Family
Qwen
Qwen
License
Tongyi Qianwen
Apache 2.0
Open weights
Yes
Yes
Release
Jun 7, 2024
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
33K
262K
Max output
8K
131K
Parameters
72B
27B
Pricing
Input $/1M
$0.35
—
Output $/1M
$0.40
—
Blended $/1M (3∶1)
$0.36
-
Speed
tok/s
65
-
TTFT (s)
0.4
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
37.1
90.5
Humanity's Last Exam
3.7
33.9
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
15.6
AA-LCR v1.1
—
82.0
CritPt
—
5.4
MMMU-Pro
—
76.3
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
61.7
SWE-bench Multilingual
—
—
LiveCodeBench
—
90.3
Terminal-Bench 2.1
—
79.8
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
—
46.6
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
42.3
DeepSWE
—
42.2
WebDev Arena
—
1,593
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,262
1,437
| Metric | Qwen2 72B Instruct | Qwen3.8 27B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Alibaba | Alibaba | — |
| Family | Qwen | Qwen | — |
| License | Tongyi Qianwen | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Jun 7, 2024 | 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 | 33K | 262K | B +229K |
| Max output | 8K | 131K | B +123K |
| Parameters | 72B | 27B | A +45 |
| Pricing | |||
| Input $/1M | $0.35 | — | — |
| Output $/1M | $0.40 | — | — |
| Blended $/1M (3∶1) | $0.36 | - | — |
| Speed | |||
| tok/s | 65 | - | — |
| TTFT (s) | 0.4 | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 37.1 | 90.5 | B +53.4 pts |
| Humanity's Last Exam | 3.7 | 33.9 | B +30.2 pts |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 15.6 | — |
| AA-LCR v1.1 | — | 82.0 | — |
| CritPt | — | 5.4 | — |
| MMMU-Pro | — | 76.3 | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | 61.7 | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 90.3 | — |
| Terminal-Bench 2.1 | — | 79.8 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | — | 46.6 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | 42.3 | — |
| DeepSWE | — | 42.2 | — |
| WebDev Arena | — | 1,593 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,262 | 1,437 | B +176 Elo |
Qwen2 72B Instruct
- Context
- 33K / 8K out
- Parameters
- 72B
- Price
- $0.35 / $0.40
- Speed
- 65 tok/s · 0.4s TTFT
- Modalities
- text → text
- License
- Tongyi Qianwen
Prior-generation Qwen2 72B instruct — predecessor to Qwen2.5 still seen in older hosted stacks.
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.