Skip to content
Blog

The State of Open-Source LLMs in 2026

Open weights route most production tokens now, Qwen leads the Hub, and China out-downloads the West. Here is the data-driven 2026 map for choosing an open model.

Published on • August 8, 2026

AI Assistant

For a long time the open-source-vs-closed debate was about potential — “can open models catch up?” In 2026 that question is closed. The Mozilla State of Open Source AI report measured the frontier capability gap at ~3%, and OpenRouter data shows open weights now route a majority of production tokens, with the seven highest-volume models on the platform all carrying open weights. The practical question is no longer “open or closed?” — it’s “which open model, and are the other advantages distributed the way you think they are?”

Open weights are where the work happens

  • Token share: from negligible at end of 2024, open-weight models went from a third of OpenRouter tokens in late 2025 to a majority by mid-2026.
  • Capability: the strongest open model scored 57 on the Artificial Analysis Intelligence Index vs. 61 for the strongest closed — fourth overall, ahead of models from three of the biggest closed labs.
  • Cost: inference prices fell up to 50× in under three years; creating an open model at ~$0.40/1M tokens is now routine.

But there’s a hard catch in the same report: open models power roughly a third of real-world AI usage while capturing only ~4% of revenue. The value is real; the business model around it is not settled yet.

The Hub has flattened geopolitically

Hugging Face grew to 13M users, 2M+ public models, and 500K public datasets. Two data points matter for anyone choosing a base:

  • China overtook the US in monthly and lifetime downloads. Chinese models were ~41% of downloads in 2025; on the top-1K, Eastern orgs (Qwen, DeepSeek, BAAI, THUDM, etc.) now command ~24% of 2.16B total downloads.
  • Qwen is the single most-downloaded org on the Hub — 399M downloads across its top models (18.5% of top-1K traffic), about 6× Meta’s Llama footprint. Qwen and DeepSeek anchored the rapid-release model cadence of 2025–26, and DeepSeek-R1 was the most-liked family on the Hub.

Apache 2.0 won the licensing war

71.5% of the top-1K models ship under a permissive open license; Apache 2.0 alone is 49%. The Llama bespoke license is still visible but plainly second-tier — fiddly derivative-use terms are an adoption tax. For serious downstream work in 2026, default to Apache 2.0 bases.

Mid-size is the production sweet spot

The “small-models-are-the-future” line is true, but blurry:

  • By count: ~51% of size-named models are below 7B.
  • By fine-tune energy: the biggest band of distinct fine-tunes in the Hub is 13–33B — where most people actually fine-tune bases.
  • By deployment economics: sub-7B wins when latency is the binding constraint; 70B+ wins when you have the GPU budget and need a frontier-ish generalist.

The mean downloaded model size rose from 827M params (2023) to 20.8B (2025) — driven by quantization and MoE weights (e.g., DeepSeek-V4 1.5T-params-MoE running at much lower active cost).

Hardware is no longer an excuse

Ways out:

  • Quantization/triple precision: BitNet-style 1.58-bit ternary weights get massive models onto consumer GPUs; 4-bit GGUF remains the pragmatic default.
  • AMD is a first-class target now (Stability, AMD-optimized kernels via the Kernel Hub).
  • Local inference on a laptop is normal; Ollama + llama.cpp self-host for eval spikes and privacy.

Choosing in 2026: a practical heuristic

  1. Default to Apache 2.0 bases. Covers half the top-1K, zero legal-review tax for derivative distributions.
  2. Evaluate by model, not lab-of-origin. South Korea and the US trend; don’t skip Qwen/DeepSeek just because the label is CN — they are the volume leaders.
  3. Match the size band to the constraint: sub-7B when latency is binding, 13–33B as the default production sweet spot, 70B+ for the frontier.
  4. Open the harness, not just the weights. The frontier has moved to the agentic harness — the orchestration loop above the model. Most open models still lack a first-party harness, and whichever got one is now sitting at the top tier of Terminal-Bench. That’s the new lock-in to watch.

Conclusion & Next Steps

Pick open for control, cost, and sovereignty; pick specific open models per task, starting Apache-2.0 mid-size unless latency is the constraint. Then evaluate the harness layer — because in 2026 the model is ~3% behind, but the software you put around it decides what it can do. Next: benchmark a Qwen-family base against a local Llama on your eval suite (not a leaderboard), compare per-request cost at your access pattern, and test auto-rotation of weights with quantized MoE checkpoints.

References / Sources

ai open-source llms hugging-face models