The open-source AI coding landscape has three clear leaders: GLM-5.1 from Z.ai, DeepSeek V3 from DeepSeek, and Qwen 3.5 from Alibaba. All three are MoE models, all are competitive with proprietary alternatives, and all are available for self-hosting.
Here’s how they compare.
Quick comparison
| GLM-5.1 | DeepSeek V3.2 | Qwen 3.5 Max | |
|---|---|---|---|
| Total params | 754B | 671B | 400B+ |
| Active params | 40B | 37B | ~50B |
| Architecture | MoE (256 experts) | MoE | MoE |
| Context | 200K | 128K | 128K |
| License | MIT | MIT | Apache 2.0 |
| SWE-Bench Pro | 58.4 | ~54 | ~52 |
| Training hardware | Huawei Ascend | NVIDIA | NVIDIA |
| Specialty | Agentic coding | Reasoning + coding | General + coding |
Coding performance
GLM-5.1 leads on SWE-Bench Pro at 58.4, which tests the hardest multi-file engineering tasks. DeepSeek V3 is strong on reasoning-heavy coding problems, and Qwen 3.5 is the most versatile — good at coding but also excellent at general tasks.
For pure coding ability, the ranking is: GLM-5.1 > DeepSeek V3 > Qwen 3.5.
But “coding” isn’t one thing. Here’s how they break down by task:
Complex multi-file refactors: GLM-5.1 wins. Its 8-hour autonomous session capability and goal alignment over thousands of tool calls make it the best choice for large-scale engineering work.
Algorithmic reasoning: DeepSeek V3 wins. DeepSeek’s reasoning models (R1, V3) are consistently strong on math and logic-heavy coding problems. If your work involves complex algorithms, data structures, or mathematical optimization, DeepSeek is the pick.
Breadth of languages and frameworks: Qwen 3.5 wins. Alibaba’s training data is the most diverse, and Qwen handles a wider range of programming languages and frameworks than the other two. It’s also the most popular model on OpenRouter by token volume.
Architecture differences
All three use Mixture-of-Experts, but the implementations differ:
GLM-5.1 uses 256 experts with 8 activated per token and DeepSeek Sparse Attention (DSA) for long-context efficiency. The 200K context window is the largest of the three.
DeepSeek V3 pioneered many of the MoE techniques that others now use. Its architecture is well-documented in their technical report and has been influential across the industry.
Qwen 3.5 uses a more compact MoE design. With fewer total parameters but more active per token, it’s often faster at inference while maintaining competitive quality.
Pricing
All three are MIT or Apache licensed, so self-hosting is free. API pricing:
| Input (per 1M tokens) | Output (per 1M tokens) | |
|---|---|---|
| GLM-5.1 (Z.ai) | ~$1.00 | ~$2.30 |
| GLM-5.1 (Coding Plan) | $3-10/month flat | Included |
| DeepSeek V3 | ~$0.27 | ~$1.10 |
| Qwen 3.5 | ~$0.30 | ~$0.60 |
DeepSeek and Qwen are significantly cheaper per token. GLM-5.1’s Coding Plan offers good value for heavy users, but for light usage, DeepSeek and Qwen win on price.
Self-hosting requirements
None of these run on consumer hardware at full precision:
| Full precision | 4-bit quantized | |
|---|---|---|
| GLM-5.1 (754B) | ~1.5TB | ~200GB |
| DeepSeek V3 (671B) | ~1.3TB | ~180GB |
| Qwen 3.5 (400B) | ~800GB | ~110GB |
Qwen 3.5 is the most practical for self-hosting due to its smaller size. All three have smaller variants available if you need something that fits on consumer GPUs.
For local development, consider the smaller models in each family: GLM-5-Turbo, DeepSeek-Coder, or Qwen 3.5 Coder.
Tool integration
GLM-5.1 has the best Claude Code integration thanks to its Anthropic-compatible API. It also works with OpenClaw, Cline, and other OpenAI-compatible tools. See our Claude Code setup guide.
DeepSeek V3 works well with most AI coding tools through its OpenAI-compatible API. It’s a popular choice for Codex CLI users.
Qwen 3.5 is available on OpenRouter (where it’s the #1 model by usage) and through Alibaba’s DashScope API. Good integration with most tools.
Which should you pick?
Pick GLM-5.1 if: You need the absolute best coding performance, especially for long-running autonomous tasks. You’re willing to pay slightly more or self-host for the best SWE-Bench scores.
Pick DeepSeek V3 if: You want the best price-to-performance ratio. DeepSeek is the cheapest option with strong coding and reasoning capabilities. Great for teams watching costs.
Pick Qwen 3.5 if: You need a versatile model that handles coding plus other tasks (writing, analysis, translation). It’s the most popular for a reason — it’s good at everything and cheap to run.
Or use all three. The beauty of open-source models is that you’re not locked in. Use GLM-5.1 for complex engineering, DeepSeek for reasoning-heavy tasks, and Qwen for everything else.
Related: GLM-5.1 Complete Guide · GLM-5.1 vs Claude vs GPT-5 · Best Free AI APIs 2026