Open-weight vs closed: what the difference actually means

A closed model (GPT-5.5, Claude, Gemini) lives behind an API — you send tokens and pay per token, and you never see the weights. An open-weight model publishes its parameters, so you can download it and run it yourself. That distinction drives almost every practical tradeoff between the two approaches.

Open-weight is not the same as open-source: Most 'open' LLMs release weights under their own licenses, not classic open-source ones. Llama uses a community license with some restrictions; DeepSeek and many Chinese models use the permissive MIT or Apache 2.0 licenses. Always check the license before commercial use.

The Llama lineage: 1 to 4

Llama's arc mirrors the field: dense models early, then a shift to Mixture-of-Experts and very long context with Llama 4. Its biggest impact, though, has been ecosystem-wide — releasing capable weights for free pushed every other lab to compete on openness and price.

The open-source ecosystem in 2026

By mid-2026 the open frontier is crowded and genuinely competitive, with several models trading blows with closed systems on coding and reasoning:

• DeepSeek V4 — 1.6T MoE, 1M context, MIT-licensed, frontier-class coding at very low cost
• Qwen (Alibaba) — broad family of open and proprietary models, strong multilingual and coding
• GLM (Z.ai) — GLM 5.2 is MIT-licensed with 1M context and top-tier open coding scores
• Mistral — Mistral Large 3 is an Apache 2.0, 675B MoE model from France
• Gemma (Google), Kimi (Moonshot), and MiMo (Xiaomi) — additional capable open releases

How to run open models

Running an open model locally usually comes down to three choices: which serving tool, how much to quantize, and how much hardware you have. Quantization shrinks the weights (for example from 16-bit to 4-bit) so a model fits in less memory, trading a small amount of quality for a large drop in resource needs.

• Ollama — simplest local runner; one command to download and chat with a model
• LM Studio / Jan — desktop apps with a graphical interface for local models
• vLLM / SGLang — high-throughput servers for production and multi-user serving
• llama.cpp — runs quantized models efficiently on CPUs and consumer GPUs

Rule of thumb: A 4-bit quantized model needs roughly half a gigabyte of memory per billion parameters. A 7-8B model runs on most modern laptops; a 70B model needs a high-end GPU or heavy quantization; frontier MoE models in the hundreds of billions of parameters generally need a multi-GPU server.

The open frontier in 2026: benchmarks and licenses

Check the license, not just the weights: 'Open' covers a range. MIT and Apache 2.0 (DeepSeek, GLM, Mistral, Qwen variants) allow unrestricted commercial use. Meta's Llama Community License permits commercial use but adds conditions. Always read the specific license before deploying.

Frequently asked questions

Is Llama free for commercial use? Mostly — the Llama Community License permits commercial use but adds some conditions (such as restrictions for the very largest platforms). For unrestricted terms, MIT or Apache 2.0 models like DeepSeek, GLM, or Mistral are cleaner choices.

Can open models match GPT-5.5 or Claude? On many coding and reasoning benchmarks, yes — DeepSeek V4 and GLM 5.2 trade blows with closed frontier models. The closed leaders still tend to edge ahead on the hardest general-reasoning tasks and on ecosystem maturity.

How do I run an open model locally? Use Ollama for the simplest setup, LM Studio or Jan for a desktop interface, or vLLM/SGLang for production serving. Quantization (e.g. 4-bit) shrinks memory needs so smaller models run on a laptop.

What is the difference between open-weight and open-source? Open-weight means the trained parameters are downloadable; open-source traditionally also implies an OSI-approved license and (ideally) training details. Most 'open' LLMs are open-weight under custom or permissive licenses.