Open-source AI, explained

What "open" actually means in AI

Three things get called "open" and only one of them matches the original open-source definition.

Open weights means the trained model file is published — you can download it, run it on your own hardware, and fine-tune it. Training code, training data, and evaluation scripts are usually notreleased. This is what almost every "open-source" model from a major lab actually is. Llama, Qwen, DeepSeek, Mistral, Gemma, Kimi K2.6, and OpenAI's gpt-oss all fall here.

Open source, in the strict sense, means weights + training code + training data + a permissive license. Very few labs do this. Allen AI's OLMo and EleutherAI's Pythia are the canonical examples — useful mostly for research, not competitive on capability with the open-weights flagships.

Open tools / open agents is the third bucket: the application layer. Source code for the program that uses an LLM to do work — agents, IDEs, orchestrators. OpenClaw, Cline, Aider, OpenHands, n8n. The model they call out to is usually closed.

The three flavors of open

Models that matter (May 2026)

Five families ship the open-weight flagships you'll actually pick from. Almost every one is a sparse Mixture-of-Experts — total parameters are huge, but only a small slice activates per token, so inference cost stays reasonable.

The leaderboard moves every month. For live rankings, check the Artificial Analysis leaderboard or LMArena.

Where to run them

Three paths, ordered by setup cost. Pick based on volume, privacy needs, and how comfortable you are with GPU plumbing.

# macOS
brew install ollama

# Pull + chat with gpt-oss-20b
# (16 GB RAM is enough)
ollama run gpt-oss:20b

# Or run DeepSeek V4 Flash if your
# GPU has the headroom
ollama run deepseek-v4:flash

curl https://api.together.xyz/v1/chat/completions \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V4-Pro",
    "messages": [
      {"role":"user","content":"hi"}
    ]
  }'

• When you're prototyping → Ollama. No account, zero cost, runs offline.
• When you ship to real users → Together or OpenRouter. Pay-per-token, OpenAI-compatible API, no GPU ops.
• When the API bill clears $10k/mo → revisit self-hosted vLLM. Below that threshold, the ops cost dwarfs the savings.

Open-source AI agents and tools

The application layer is where most of the practical action is in 2026. These tools are open-source — you can read every prompt, every tool definition, every loop — even when the model behind them is closed.

For the full list and head-to-head comparisons, see /tools and /compare.

Open vs frontier — when each wins

Common pitfalls

• Treating "open" as one license.Llama's license bans use by anyone with >700M MAU and includes acceptable-use clauses. Apache 2.0 (Qwen, Gemma, Mistral, gpt-oss) is genuinely permissive. MIT (DeepSeek) is the most generous. Read the license before you build a product on the model.
• Assuming open-weights is always behind frontier. Not true for coding in 2026 — Kimi K2.6 ties GPT-5.5 on SWE-Bench Pro, DeepSeek V4 Pro leads LiveCodeBench. Frontier still wins on the hardest reasoning benchmarks and on tool-use reliability, but the gap is task-specific, not uniform.
• "Free" means you pay the GPU bill.Self-hosting a 400B-active model isn't free — it's a $30k+ box plus power. Managed inference (Together, Groq, Fireworks) is cheaper than the API-equivalent frontier model but still costs real money per token. Compare token rates honestly.
• Quantization quality loss is real.A 4-bit quant of a 70B model isn't the same model — capability degrades, especially at long context and on reasoning. For benchmarks, run the quant you'll deploy, not the full-precision reference.
• Open agents inherit prompt-injection risk. An agent like OpenClaw with broad permissions (inbox, terminal, calendar) is exactly the high-blast-radius surface attackers target. Sandbox aggressively, audit any third-party skills, and read what you install.
• Confusing inference-cheap with training-cheap.MoE architectures keep inference affordable, but pre-training a frontier-class open-weights model still costs tens of millions of dollars. The reason you're getting these for free is that someone else paid for training and chose to release the weights — not because LLMs got cheap to build.

Terms worth knowing