Qwen3 VL 32B (Reasoning) vs Qwen3.8 27B
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Qwen3 VL 32B (Reasoning)10.1
Qwen 27B14.1
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Qwen3 VL 32B (Reasoning)—
Qwen 27B5.1
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Qwen3 VL 32B (Reasoning)—
Qwen 27B40.3
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
Qwen3 VL 32B (Reasoning)—
Qwen 27B1,366
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
Qwen3 VL 32B (Reasoning)63.4
Qwen 27B74.2
| Benchmark | Qwen3 VL 32B (Reasoning) | Qwen 27B |
|---|---|---|
| Humanity's Last Exam | 10.1 | 14.1 |
| Terminal-Bench 4.0 | — | 5.1 |
| AutomationBench-AA | — | 40.3 |
| Arena Text · creative writing | — | 1,366 |
| MMMU-Pro (vision) | 63.4 | 74.2 |
Specs
| Spec | Qwen3 VL 32B (Reasoning) | Qwen 27B |
|---|---|---|
| Input, per 1M | $0.20 | $0.42 |
| Output, per 1M | $1.50 | $2.55 |
| Cost per task? | — | $1.13 |
| Speed, tok/s | — | 50 |
| Context | 256K | 262K |
| Weights | Open | Open |
| Released | Oct 2025 | Aug 2026 |
| Memory (4-bit) | 20 GB | 17 GB |
| Licence | Apache-2.0 | Apache-2.0 |
| Commercial use | Allowed | Allowed |
| See images | ✓ | ✓ |
| Read PDFs | ? | Partial |
| Hear audio | — | — |
| Watch video | — | ✓ |
| Call tools | ✓ | ✓ |
| Force a tool call | ? | Partial |
| Return strict JSON | Partial | ✓ |
| Search the web | ? | ? |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
Qwen3 VL 32B (Reasoning) → Qwen 27B
max_tokens range · QwenCloud›
With thinking enabled, max_tokens outside 1–32768 returns HTTP 400; use max_completion_tokens to budget reasoning plus answer without that cap.
Static YaRN affects shorter texts · Self-hosted›
Static YaRN can hurt short-text performance; change rope_parameters only when longer contexts are required.
tool_choice only supports auto or none · QwenCloud›
With thinking enabled, tool_choice supports only auto or none.
Include the word JSON · QwenCloud›
JSON Object mode requires the word JSON in a system or user message; otherwise the API returns an error.
Agent search_strategy not supported · QwenCloud / Chat Completions›
Chat Completions does not support Qwen3.8 agent search_strategy; use the Responses web_search tool for multi-turn retrieval.
Qwen 27B → Qwen3 VL 32B (Reasoning)