1.2 Which Models Are Suitable for Coding? A Cheat Sheet for Model Types: Instruct, Thinking, Densed, Multi-modal, MoE, MTP, MLX, Quantization

I kept talking one of my colleagues’ ears off about trying an LLM to solve a coding task for a fairly large feature. Eventually, he gave in and tried it, after which he spent a long time complaining and wondering how stupid the neural network was. I hadn’t explained or told him anything, thinking he would figure it out himself. As it turned out, he had used the ChatGPT 4-mini model for coding, which is a very small and stupid model for tasks like these. That’s why I decided this article would be worthwhile.

It’s easy to get confused in the world of LLMs—there are so many of them now. Even if you use subscriptions, you still need to understand what each model is intended for and which model will be sufficient for a particular task. Of course, you can use the smartest model for everything—but that isn’t efficient. Top-tier, heavyweight models usually work 2–5 times slower than their younger siblings. For translating ordinary text, there’s no point in using a top-tier model: an average one will do just as well, but three times faster and at a significantly lower cost. It’s like making a PhD solve a schoolchild’s multiplication table.

When it comes to local models, you really need to understand the full range of options, since you’ll have to find the best models for the hardware available to you. This is where you’ll truly feel the difference—for example, between two 27-billion-parameter models, one trained almost entirely on someone else’s code and the other a general-purpose model with built-in Vision. When it comes to writing code, the difference between them will be astronomical.

What kinds of models are there, anyway?

Models can be very different, and I want to tell you about the main things you need to know.

By purpose and operating logic (Interaction Types)

• Instruct (Instruction-tuned)

  • Fine-tuned to understand user commands.

  • A rarity: obsolete and unsuitable for everyday tasks, but they can be useful for automation on low-end hardware and may be faster because they do not have a reasoning module. Still, newer models will be better at everything anyway.

  • They undergo instruction tuning (SFT) and alignment (RLHF).

  • Example: Models that respond to direct prompts rather than simply continuing the text.

• Thinking / Reasoning

  • Equipped with a hidden “thinking” stage before producing an answer.

  • Use chains of thought (Chain-of-Thought) to solve complex logical problems.

  • Example: Qwen 27B, DeepSeek-R1, all top-tier models

• Multi-modal

  • Work with several types of data simultaneously.

  • Accept or generate text, images, audio, and video.

  • Example: GPT-4o, Gemini 1.5, all top-tier models

  • The screenshot shows an example of a small Qwen 3.5 9B model that can recognize images, think, and call tools.

    

By architecture and weight structure

• Dense

  • The traditional neural network architecture. For coding, it offers deeper understanding than an MoE model of the same size.

  • Each token (word) activates 100% of the model's parameters.

  • Example: Standard Llama 3 or Qwen 3.6 27B models.

• MoE (Mixture of Experts)

  • The model is divided into subnetworks (“experts”).

  • For each token, the router activates only a subset of the experts (for example, 2 out of 8).

  • This saves computational resources while maintaining high quality.

  • Examples: Qwen3.6-35B-A3B, Mixtral 8x7B, DeepSeek-V3.

By optimization methods and formats

• MTP (Multi-Token Prediction)

  • An architectural approach in which the model predicts not just the next token, but several tokens at once.

  • It speeds up text generation and improves the understanding of long-term context.

  • A 20–200% speed increase, depending on the complexity and routine nature of the task

  • Example: Qwen 3.6 27B MTP

• Quantization

  • Compressing the model by reducing the precision of its numerical weights (for example, from 16-bit to 8-bit, 4-bit, or even 1.58-bit). For consumer hardware, a format such as Q4_K_M is usually chosen, as it is generally the most well-balanced option. For example, Q8_0 involves almost no compression and virtually no quality loss, but requires significantly more VRAM.

  • A model is a set of floating-point numbers. Quantization usually reduces the number of digits after the decimal point: for example, 0.3456783945673456 becomes 0.34567839, reducing the number of decimal places from 16 to 8.

  • Reduces the model size and video memory (VRAM) requirements.

  • Makes it possible to run large models on home PCs and smartphones.

    

• MLX

  • A specialized framework from Apple for efficient model training and inference. It provides a 20–40% increase in token generation speed.

  • Optimized for Apple Silicon chips (M1/M2/M3/M4).

Models in the MLX format use the Mac's Unified Memory at maximum speed.

Cloud-based coding models available by subscription: Anthropic (Claude Code), ChatGPT (Codex)

I won't go into detail here. These top-tier models are also used to create new versions of themselves. They are universal, suitable for coding tasks as well as any other tasks.

I want to show you a table using ChatGPT as an example—pay attention to the speed.

Comparison of models in the GPT-5.6 lineup (31.07.2026)

The picture is comparable for Anthropic (Claude Code).

Local or enterprise coding models: Qwen 3.6 27B and Qwen 3.6-35B-A3B

Here's a small comparison table.

Comparison of Qwen 3.6 27B and Qwen 3.6-35B-A3B

For local coding, Qwen is the favorite. But you can also consider DeepSeek and GLM.

I won’t overload you with unnecessary information—this will be enough for now.