vs
9:2
overall capability wins (1 ties across 12 scored). Qwen3.8 27B leads overall.
- Reasoning3:1
- Coding3:0
- Arena1:0
- Tool use1:0
- Specs1:1
Verdict
Qwen3.8 27B leads coding (widest gap: +73.1 on TermBench); Qwen3.8 27B ranks higher on LMArena (+38 Elo); Qwen3.8 27B allows ~8.0x more max output (131K vs 16K).
- Qwen3.8 27B leads coding (widest gap: +73.1 on TermBench)
- Qwen3.8 27B ranks higher on LMArena (+38 Elo)
- Qwen3.8 27B allows ~8.0x more max output (131K vs 16K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Qwen3.8 27B
- -
- Qwen3-Next 80B-A3B
- $0.29
Price per 1M tokens
Full comparison
Organization
Alibaba
Alibaba
Family
Qwen
Qwen
License
Apache 2.0
Apache 2.0
Open weights
Yes
Yes
Release
Aug 14, 2026
Sep 15, 2025
Knowledge cutoff
-
-
API / provider
No primary API price listed
Alibaba / Together (ref.)
Modalities
text, image, video → text
text → text
Specs
Context window
262K
262K
Max output
131K
16K
Parameters
27B
80B (3B act.)
Pricing
Input $/1M
—
$0.09
Output $/1M
—
$0.78
Blended $/1M (3∶1)
-
$0.26
Speed
tok/s
-
110
TTFT (s)
-
0.3
Reasoning
MMLU-Pro
—
80.6
GPQA Diamond
90.5
75.9
Humanity's Last Exam
33.9
12.6
AIME 2025
—
69.5
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
15.6
19.0
AA-LCR v1.1
82.0
63.7
CritPt
5.4
—
MMMU-Pro
76.3
—
IFBench
—
60.7
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
61.7
—
SWE-bench Multilingual
—
—
LiveCodeBench
90.3
56.6
Terminal-Bench 2.1
79.8
6.7
Aider Polyglot
—
49.8
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
46.6
38.8
CursorBench
—
—
SWE-Rebench
—
20.1
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
1,437
1,399
| Metric | Qwen3.8 27B | Qwen3-Next 80B-A3B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Alibaba | Alibaba | — |
| Family | Qwen | Qwen | — |
| License | Apache 2.0 | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Aug 14, 2026 | Sep 15, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | Alibaba / Together (ref.) | — |
| Modalities | text, image, video → text | text → text | — |
| Specs | |||
| Context window | 262K | 262K | tie |
| Max output | 131K | 16K | A +115K |
| Parameters | 27B | 80B (3B act.) | B +53 |
| Pricing | |||
| Input $/1M | — | $0.09 | — |
| Output $/1M | — | $0.78 | — |
| Blended $/1M (3∶1) | - | $0.26 | — |
| Speed | |||
| tok/s | - | 110 | — |
| TTFT (s) | - | 0.3 | — |
| Reasoning | |||
| MMLU-Pro | — | 80.6 | — |
| GPQA Diamond | 90.5 | 75.9 | A +14.6 pts |
| Humanity's Last Exam | 33.9 | 12.6 | A +21.3 pts |
| AIME 2025 | — | 69.5 | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 15.6 | 19.0 | B +3.4 pts |
| AA-LCR v1.1 | 82.0 | 63.7 | A +18.3 pts |
| CritPt | 5.4 | — | — |
| MMMU-Pro | 76.3 | — | — |
| IFBench | — | 60.7 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | 61.7 | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | 90.3 | 56.6 | A +33.7 pts |
| Terminal-Bench 2.1 | 79.8 | 6.7 | A +73.1 pts |
| Aider Polyglot | — | 49.8 | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 46.6 | 38.8 | A +7.8 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | 20.1 | — |
| 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 | 1,437 | 1,399 | A +38 Elo |
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).
Qwen3-Next 80B-A3B
- Context
- 262K / 16K out
- Parameters
- 80B (3B act.)
- Price
- $0.09 / $0.78
- Speed
- 110 tok/s · 0.3s TTFT
- Modalities
- text → text
- License
- Apache 2.0
Ultra-sparse 80B MoE (3B active) — high throughput instruct model in the Qwen3 family.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.