Skip to main content
  1. Posts/

Open models for programming

··462 words·3 mins·

🚀 7 open‑source models that are changing AI coding
#

These are open‑source AI models for coding without relying on cloud services.
The core idea is simple: more privacy, more control, and zero API costs.

Key takeaways:

  • 🔐 Full privacy: your code never leaves your machine.
  • Powerful models: from advanced reasoning to autonomous agents.
  • 🧩 MoE and huge contexts: perfect for long workflows, debugging, and complex tasks.
  • 💸 Savings: if you already have good hardware, you can avoid expensive subscriptions.

🧠 TL;DR
#

Imagine that instead of sending your code to external servers (like when using Copilot or Claude), you could run a “mini‑AI brain” locally on your computer.
That means:

  • No one sees your code.
  • Fast responses without depending on the internet.
  • You can automate long tasks without limits.

These open‑source models enable exactly that: high‑level AI, but on your own machine.

🧩 Models and their main features
#

1. Kimi‑K2‑Thinking (Moonshot AI)
#

  • 🧠 1T‑parameter MoE (32B active)
  • 🛠️ Agent with step‑by‑step reasoning
  • 🔁 Maintains 200–300 tool calls without losing coherence
  • 📏 Context: 256K tokens
  • INT4 optimized for low latency
  • ⭐ Strong in long reasoning, multilingual, and autonomous workflows

2. MiniMax‑M2 (MiniMaxAI)
#

  • 🧩 230B‑parameter MoE (10B active)
  • High efficiency and low latency
  • 🔄 Ideal for plan → act → verify loops
  • 🎯 Built for interactive agents and coding tasks
  • 💰 Optimized for cost and speed

3. GPT‑OSS‑120B (OpenAI)
#

  • 🧠 117B parameters, 5.1B active
  • 🔧 Native tools: function calling, browsing, Python, structured outputs
  • 🎚️ Configurable reasoning levels
  • 🧪 Full fine‑tuning available
  • 🥇 High performance in benchmarks, reasoning, and tool use

4. DeepSeek‑V3.2‑Exp (DeepSeek AI)
#

  • 🧠 671B parameters, 37B active
  • 🧵 Introduces DeepSeek Sparse Attention (DSA)
  • 📏 128K token context
  • 🎯 Optimized for long‑sequence efficiency
  • 🔬 Performance similar to V3.1 but with efficiency gains

5. GLM‑4.6 (Z.ai)
#

  • 🧠 355B parameters, 32B active
  • 📏 Extended context to 200K tokens
  • 💻 Clear improvements in coding and frontend generation
  • 🔧 Better integration with agents and tools
  • 🥇 Competitive against DeepSeek‑V3.1 and Claude Sonnet 4

6. Qwen3‑235B‑A22B‑Instruct‑2507 (Alibaba Cloud)
#

  • 🧠 235B parameters, 256K tokens
  • 🎯 No‑thinking model: direct answers without showing reasoning
  • 🌍 Strong in multilingual, logic, math and coding
  • 🧰 Improvements in tool use and alignment with user preferences
  • 🏢 Ideal for practical, production tasks

7. Apriel‑1.5‑15B‑Thinker (ServiceNow AI)
#

  • 🧠 15B parameters (very compact)
  • 👁️ Multimodal: text + images
  • 📏 ~131K token context
  • ⚙️ Continuous training on text and images
  • 🏭 Excellent for enterprise agents and DevOps

More information at the link 👇

Also published on LinkedIn.
Juan Pedro Bretti Mandarano
Author
Juan Pedro Bretti Mandarano