Open-source coding models now match or beat proprietary alternatives. Here’s the definitive ranking for April 2026, based on benchmarks, real-world performance, and practical usability.

Update (April 24, 2026): DeepSeek V4-Pro (1.6T/49B active, MIT license) is the new #1 open-source coding model — 80.6% SWE-bench Verified, Codeforces 3206, 93.5% LiveCodeBench. See the V4 Pro guide.

Update (April 23, 2026): Alibaba released Qwen 3.6-27B, a 27B dense model scoring 77.2% on SWE-bench Verified — beating the 397B flagship. Apache 2.0 licensed, runs on 22GB VRAM. It slots in as a top-tier open-source coding model.

The ranking

1. DeepSeek V4-Pro — Best overall

• Parameters: 1.6T MoE (49B active)
• SWE-Bench Verified: 80.6%
• Codeforces: 3206
• LiveCodeBench: 93.5%
• License: MIT
• Best for: Frontier-class coding, competitive programming, autonomous agents
• Limitation: Requires enterprise hardware or API access for self-hosting

DeepSeek V4-Pro takes the #1 spot with 80.6% on SWE-bench Verified, Codeforces rating of 3206, and 93.5% on LiveCodeBench. It has 1M context, MIT license, and API pricing at $1.74/$3.48 per 1M tokens. For on-demand access to enterprise GPUs, see cloud GPU providers. The V4-Flash variant (284B/13B active) scores 79.0% on SWE-bench at just $0.14/$0.28.

2. GLM-5.1 (Z.ai) — Best value

• Parameters: 754B MoE (40B active)
• SWE-Bench Pro: 58.4 (#1 overall, including proprietary)
• License: MIT
• Best for: Complex multi-file engineering, autonomous coding sessions
• Limitation: Requires enterprise hardware or API access

GLM-5.1 is the first open-source model to top SWE-Bench Pro, beating GPT-5.4 and Claude Opus 4.6. Its 8-hour autonomous coding capability is unmatched. The MIT license makes it the most permissive frontier model available.

Use it via the GLM Coding Plan ($3/month) or self-host if you have the hardware.

3. DeepSeek V3.2 — Budget pick

• Parameters: 671B MoE (37B active)
• SWE-Bench Pro: ~54
• License: MIT
• Best for: Reasoning-heavy coding, algorithmic problems
• Limitation: Slightly behind GLM-5.1 on complex engineering tasks

DeepSeek V3 pioneered many MoE techniques now used across the industry. It’s the cheapest frontier-class model per token (~$0.27/1M input) and excels at mathematical and algorithmic coding. The reasoning variants (R1) are particularly strong.

4. Qwen 3.5 (Alibaba) — Most versatile

• Parameters: 400B+ MoE
• SWE-Bench Pro: ~52
• License: Apache 2.0
• Best for: General-purpose coding + other tasks
• Limitation: Not as specialized for coding as GLM-5.1

Qwen 3.5 is the #1 model on OpenRouter by token volume for a reason — it’s good at everything. Coding, writing, analysis, translation. If you want one model for all tasks, Qwen is the pick. The Coder variant is optimized for development work.

5. Gemma 4 27B (Google) — Best for local use

• Parameters: 27B dense
• License: Gemma License
• Best for: Local development, fast completions
• Limitation: Can’t match frontier models on complex tasks

Gemma 4 is the best model you can run on consumer hardware. The 27B variant fits on a single RTX 4090 or a Mac with 32GB RAM and delivers impressive coding quality for its size. Perfect for daily development.

6. Llama 4 Scout (Meta) — Best ecosystem

• Parameters: 109B MoE (17B active)
• License: Llama License
• Best for: Broad tool integration, fine-tuning
• Limitation: Meta’s shift to proprietary Muse Spark raises questions about future Llama investment

Llama 4 has the largest ecosystem of fine-tunes, tools, and community support. Scout is efficient enough to run on consumer hardware while delivering solid coding performance. The 10M token context window is the largest available.

7. GLM-5-Turbo (Z.ai) — Best speed/quality tradeoff

• Parameters: Smaller than GLM-5.1
• License: Proprietary
• Best for: Fast coding with good quality
• Limitation: Not open-source (proprietary license)

GLM-5-Turbo is Z.ai’s faster variant, optimized for lower latency. Good for interactive coding where you need quick responses. Available through the GLM Coding Plan.

8. Gemma 4 12B (Google) — Best for constrained hardware

• Parameters: 12B dense
• License: Gemma License
• Best for: Running on laptops, Raspberry Pi, edge devices
• Limitation: Limited on complex multi-file tasks

The 12B Gemma 4 runs on 8GB of VRAM — that’s an RTX 3060 or a MacBook with 16GB RAM. Surprisingly capable for its size, especially for code completions and small edits.

9. MiMo V2 Pro (Xiaomi) — Best small reasoning model

• Parameters: ~8B
• License: Apache 2.0
• Best for: Reasoning-heavy coding on limited hardware
• Limitation: Small context window, limited on large codebases

Xiaomi’s MiMo V2 Pro punches above its weight on reasoning tasks. At 8B parameters, it runs on almost any modern GPU and delivers surprisingly good results on algorithmic problems.

Comparison table

Model	Params (active)	SWE-Bench Pro	License	Min VRAM	API cost
DeepSeek V4-Pro	49B	80.6% (Verified)	MIT	200GB+	$1.74/1M
GLM-5.1	40B	58.4	MIT	200GB+	$3/mo plan
DeepSeek V3.2	37B	~54	MIT	180GB+	$0.27/1M
Qwen 3.5	~50B	~52	Apache 2.0	110GB+	$0.30/1M
Gemma 4 27B	27B	—	Gemma	16GB	Free (local)
Llama 4 Scout	17B	—	Llama	12GB	Free (local)
Gemma 4 12B	12B	—	Gemma	8GB	Free (local)
MiMo V2 Pro	~8B	—	Apache 2.0	6GB	Free (local)

How to choose

“I want the best coding AI, period” → DeepSeek V4-Pro via API or V4 Pro guide

“I want good coding AI for free” → Gemma 4 27B locally with Ollama

“I want the cheapest API” → DeepSeek V3 or Qwen 3.5

“I want one model for everything” → Qwen 3.5

“I have a laptop with 16GB RAM” → Gemma 4 12B or Llama 4 Scout

“I’m building a product” → GLM-5.1 (MIT license, best performance) or DeepSeek V3 (MIT, cheapest)

The open-source coding model landscape has never been stronger. Two years ago, you needed proprietary models for serious coding work. Today, the best coding model in the world is open-source and MIT-licensed — DeepSeek V4-Pro at 80.6% SWE-bench Verified.

FAQ

What’s the best open-source coding model in 2026?

Devstral 2 is the best open-source coding model, scoring 72.2% on SWE-bench Verified with 256K context and MIT license. It handles multi-file edits, understands type systems, and works with all major coding tools like Aider and Continue.dev.

Can open-source coding models replace GitHub Copilot?

Yes. Running Codestral 22B locally with Continue.dev gives you equivalent autocomplete quality. For chat-based coding assistance, Qwen 2.5 Coder 32B or Devstral 2 match Copilot’s capabilities while keeping your code completely private.

Which open-source coding model runs on consumer hardware?

Qwen 2.5 Coder 14B runs on 8GB VRAM and handles most coding tasks well. Codestral 22B fits on 12GB VRAM and is the best for autocomplete. For 16GB VRAM, Devstral Small 24B offers near-frontier coding quality.

Related: Qwen 3.6 Plus Guide · GLM-5.1 Complete Guide · Best AI Models for Coding Locally · Best Free AI APIs 2026