Every few months someone claims open-source has “caught up.” Then you try to deploy the model — and the benchmark score doesn’t survive contact with production.
Here’s what actually works in mid-2026, tier by tier, with the honest weaknesses. In 5 minutes you’ll know which open model to pick — or when to stay on a closed API.
For a broader map, see our best free AI guide, ChatGPT vs Gemini and what is AI.
The tiers that matter
Open-weight releases now cluster into three practical tiers:
- Frontier tier (Llama-4, DeepSeek-V4, Qwen-3) — competes with GPT-5 on single-turn evals, loses gracefully on agentic ones.
- Efficient tier (8–14B) — where most production deployments land.
- Tiny tier (under 3B) — finally usable for on-device.
Comparison
Here’s the short version:
| Tier | Leader | Best for | Approx. cost (hosted) |
|---|---|---|---|
| Frontier | Llama-4-Instruct | General chat, reasoning | ~$1.20 / M tokens |
| Frontier | DeepSeek-V4 | Cheapest frontier tier | ~$0.55 / M tokens |
| Efficient | Qwen-3-14B | Balanced production | ~$0.30 / M tokens |
| Tiny | Phi-4-mini | On-device, edge | ~free (local) |
What the benchmarks miss
Instruction-following on 5+ turn conversations is where closed models still win.
Here’s the catch: tool-calling reliability under noisy input is another gap. If your product depends on agentic reliability, budget for it.
The pragmatic pick
For most teams:
- General chat: Llama-4-Instruct
- Code: DeepSeek-Coder-V3
- Vision: Qwen-3-VL
Anything else is either overkill or a benchmark chase. For the beginner path, start with how to use AI.