Qwen3.5-27B leads on the shared-benchmark average
Ahead on 6 of 6 shared benchmarks, averaging 25.8 points higher
Summarised only from the 6 percentage-scale benchmarks scored by every selected model; details are below. A further 1 historical, reverse-direction or differently scaled metrics are excluded; indexes, cost and time are not added to accuracy percentages.
“Best available” takes each model’s best recorded non-parallel result in the metric’s stated direction per benchmark, so it may combine modes into a virtual configuration that does not exist. Read it with the mode breakdown.

Qwen3.5-27B
阿里巴巴
Benchmark-by-benchmark comparison. Changing the thinking mode or tool filters updates the chart and table below.
“Best available” picks the best non-parallel result in the metric’s stated direction separately for each benchmark. The resulting series can combine several reasoning levels and is not one reproducible runtime configuration. Choose a mode filter to compare like-for-like runs.
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
Each axis is the mean percentage score of one benchmark domain. It is an average, not a capability rating.
Relative edge: Agent能力评测 +45.0 / Relative gap: none clear
Relative edge: none clear / Relative gap: Agent能力评测 -45.0
Method: for each model and benchmark, all scores in the current mode scope are averaged (not the best score), then those benchmark scores are averaged within each domain. Only benchmarks scored on a 0-100 scale by at least two of the selected models count — Elo and rating-scale benchmarks such as Codeforces or Arena are excluded, because averaging a 1500 rating with an 85% accuracy produces a meaningless number. Missing values are not counted as zero, and the averages are unweighted, so domains with harder benchmarks read lower.
Every model and runtime mode, benchmark by benchmark. Values are comparable along a row, not between different benchmarks.
9 benchmarks with comparable scores. Each model shows its best score; mode label is displayed below.
| Benchmark | Qwen3.5-27B | Qwen3-32B |
|---|---|---|
90.50Thinking Enabled | 87.30Thinking Enabled | |
0.90Thinking Enabled | 0.30Thinking Enabled | |
48.50Thinking Enabled | Tools | 7.40Thinking Enabled | |
85.50Thinking Enabled | 68.40Thinking Enabled | |
1899.00Thinking Enabled | 1977.00Thinking Enabled | |
80.70Thinking Enabled | Tools | 65.70Thinking Enabled | |
32.60Thinking Enabled | Tools | 3.00Thinking Enabled | Tools | |
93.90Thinking Enabled | Tools | 29.80Thinking Enabled | Tools | |
76.50Thinking Enabled | 36.30Thinking Enabled |
Official list prices per model API, split by input and output. Unit: USD per 1M tokens.
Architecture, licensing and API modalities. "Not provided" means the field is missing from our database.
| Features & specs | Qwen3.5-27B阿里巴巴 | Qwen3-32B阿里巴巴 |
|---|---|---|
Core specsRelease | 2026-02-25 | 2025-04-28 |
Context length | 1010K | 128K |
Total parameters | 27B | 32B |
Max output length | 248,320 tokens | 16,384 tokens |
Architecture | Dense | Dense |
Runtime modes | 关闭开启扩展 | No runtime mode data |
Availability & licensingCode availability | Available · Qwen License | Available · Apache 2.0 |
Weight availability | Available · Qwen License | Available · Apache 2.0 |
Use & commercial terms | 免费商用授权 | 免费商用授权 |
Local deploymentWeight size | 55.6 GB | 64GB |
VRAM for weights | ≈ 55.6 GB (weights only, excludes KV cache) | ≈ 64 GB (weights only, excludes KV cache) |
Weights | Hugging Face | Hugging Face |
Source repo | GitHub | Not provided |
API modality supportText Input/Output | Input:YesOutput:Yes | Input:YesOutput:Yes |
Image Input/Output | Input:YesOutput:No | Input:NoOutput:No |
Video Input/Output | Input:YesOutput:No | Input:NoOutput:No |
ResourcesPaper / report | Qwen3.5: Towards Native Multimodal Agents | Qwen3: Think Deeper, Act Faster |
DataLearner blog | Not provided | 重磅!阿里开源第三代千问大模型:Qwen3系列,最小仅6亿参数规模,最大2350亿参数规模大模型! |

Qwen3-32B
阿里巴巴