Qwen3 Is Here: What's New, Benchmarks, and How to Upgrade
Alibaba's Qwen3 is the latest generation of their open-weight model family, and it's a meaningful step up. Compared to Qwen2.5 at the same parameter count, Qwen3 delivers noticeably better reasoning on math and coding tasks, stronger instruction-following, and a revamped tool-calling implementation that makes it more reliable as an agent backbone.
The headline feature is an optional "thinking" mode — the model can reason step-by-step through a problem before giving a final answer, similar to o1-style chain-of-thought, with the difference that you can turn it off when speed matters more than depth. It's available on all size variants from 4B to 235B.
What changed
- 📐 Math and reasoning
Qwen3 14B matches or exceeds Qwen2.5 32B on MATH-500 and GSM8K benchmarks — effectively getting one size tier of quality improvement for free.
- 💻 Coding performance
HumanEval scores jump from ~71% (Qwen2.5 14B) to ~83% (Qwen3 14B). The model produces more complete functions and is better at recovering from its own errors when given feedback.
- 🔧 Tool-calling reliability
The structured output format for tool calls is more consistent, reducing malformed JSON responses that caused agent loops in smaller Qwen2.5 variants. This makes Qwen3 8B+ viable for agent workloads that previously required 14B+.
- 📏 Longer context window
Context window extended to 128K tokens across all variants (up from 32K in Qwen2.5). Useful for summarizing long documents, processing codebases, or maintaining extended agent sessions without truncation.
Benchmark comparison
All figures are from Alibaba's technical report and independent community reproductions. Qwen3 14B is the primary comparison point since it's the most commonly used size.
| Benchmark | What it measures | Qwen2.5 14B | Qwen3 14B | Qwen3 32B |
|---|---|---|---|---|
| MMLU | General knowledge | 79.8 | 85.3 | 89.1 |
| HumanEval | Python coding | 71.3 | 82.9 | 90.2 |
| GSM8K | Grade-school math | 85.6 | 93.1 | 96.4 |
| MATH-500 | Competition math | 62.4 | 76.8 | 84.3 |
Scores are pass@1 accuracy (%) or accuracy (%) depending on benchmark. Higher is better. Qwen3 14B thinking-mode disabled for fair comparison with non-CoT baselines.
How to upgrade
Upgrading is a single ollama pull command.
Ollama handles the download and automatically updates the model tag to point to the new weights.
If you have Qwen2.5 installed and want to reclaim the disk space, remove it after confirming
Qwen3 works as expected:
# Pull the latest Qwen3 14B (downloads ~9 GB)
ollama pull qwen3:14b
pulling manifest... done
pulling a1b2c3d4e5f6... 100% ▕████████████████▏ 9.1 GB
verifying sha256 digest... done
status: success
# Optional: remove the old Qwen2.5 14B to free ~9 GB
ollama rm qwen2.5:14b
# Available size variants
ollama pull qwen3:4b # ~2.6 GB — fast, low-end hardware
ollama pull qwen3:8b # ~5.2 GB — good balance
ollama pull qwen3:14b # ~9.1 GB — recommended
ollama pull qwen3:32b # ~21 GB — best quality, needs 24 GB+ VRAM/think and
/no-think in the interactive CLI,
or set think: false in the API
request body for programmatic use. Thinking mode uses significantly more tokens per response —
budget 2–4× the usual token count for complex tasks.
Should you upgrade?
Yes, for most users — especially if you use your local model for coding, math, or as an agent backbone. The improvement in coding accuracy and tool-calling reliability at the 14B tier alone justifies the 9 GB re-download.
If your hardware is very constrained (under 12 GB of VRAM or RAM) and Qwen2.5 7B is already running well for your use case, there's no urgent reason to switch — Qwen2.5 remains a solid model and still receives community support. But if you're already on 14B, the upgrade is straightforward and the gains are real.
Not sure which Qwen3 size fits your hardware?
Our Model Picker guide maps RAM and VRAM amounts to specific model sizes and quantizations — so you know exactly which variant to pull before downloading 21 GB and finding out it doesn't fit.
Go to Model Picker →