Fifteen families — the open-weight LLM landscape has hardened into mature infrastructure
Llama, Kimi, DeepSeek, Qwen, Yi, Gemma, Mistral, Phi, Granite, GLM, plus five more. Fifteen-plus shipping open-weight model families, ~80-100 production-grade models, every quality tier covered. The open-weight ecosystem is no longer the experimental edge of AI — it's the mature parallel stack to the closed-frontier offering.
Twelve months ago the open-weight AI landscape was a small handful of families with shipping frontier-class models: Meta's Llama, Alibaba's Qwen, Mistral, and DeepSeek as the credible four. Other families existed but were research-grade rather than production-grade. The implicit procurement frame was "open-weight is the cost-conscious or compliance-driven alternative when closed-frontier doesn't fit."
May 2026 is structurally different. A comprehensive survey published this cycle maps 15+ active open-weight LLM families with shipping production-grade models — covering every capability tier from sub-billion-parameter edge models through trillion-parameter MoE frontier models. The ecosystem has hardened into a mature parallel stack with credible vendors at every capability point.
The 15 families and the production count
The complete set as of May 2026: Llama (Meta), Kimi (Moonshot), DeepSeek, Qwen (Alibaba), Yi (01.AI), Gemma (Google), Mistral, Phi (Microsoft), Granite (IBM), GLM (Zhipu), plus secondary families: Falcon (TII), OLMo (AI2), RedPajama (Together), Pythia (EleutherAI), plus the various academic and corporate spin-offs.
Each family has multiple model sizes. Mistral alone ships Small 4, Medium 3.5, Large 3, plus specialized Voxtral TTS, Leanstral, Forge, and Ministral variants. Qwen has 3.5 / 3.6 / 3.7 / 3.7 Max plus the 27B and 235B-A22B sizes. DeepSeek ships V4 Pro and V4 Flash. IBM's Granite 4.1 8B New just launched joining the existing Granite family. The total count of shipping production-grade open-weight models across all 15+ families is approximately 80-100 — large enough to make per-task model selection a genuine optimization problem rather than a vendor-choice problem.
Why the hardening matters for procurement
Mature parallel stacks have three procurement properties that experimental edges don't have. First: vendor risk is distributed. Losing any one open-weight provider doesn't break the procurement strategy because alternatives exist at every quality tier. If Meta stops shipping Llama (which is increasingly plausible given the 13-month silence since Llama 4), procurement teams can replace it with Mistral, Qwen, DeepSeek, or several others without changing the overall architecture.
Second: architectural convergence simplifies tooling. Every flagship 2026 open-weight model is sparse MoE. The deployment infrastructure (vLLM, TGI, Ollama, SGLang) works across the entire ecosystem with minimal per-vendor specialization. Procurement teams don't need to learn a new deployment stack per vendor.
Third: price-performance is dense along the Pareto frontier. There's a credible option at every capability-and-cost point. Procurement teams can run cost-quality optimization as a continuous tuning problem rather than a binary vendor-choice problem. That's the level of market maturity that lets enterprises spend AI procurement dollars efficiently rather than just spending them.
The Meta question still hangs
Llama 4 shipped in April 2025. Llama 5 has not appeared in the 13 months since. If the silence continues into 2026 H2, the Western open-weight frontier is structurally a Mistral-plus-secondary-tier story — with the Chinese cohort (Qwen, DeepSeek, Moonshot Kimi, Zhipu GLM, Yi) dominating frontier-class open-weight quality.
For procurement teams with compliance constraints on Chinese-origin open weights (US federal customers, regulated-industry buyers with export-control exposure), Mistral plus Granite plus Phi is the deployable cohort — three Western-domiciled families. That's a meaningfully smaller stack than the full 15-family ecosystem, but it's still large enough to support most procurement scenarios. The federal-procurement-only constraint costs roughly an order of magnitude in available model count, which translates to slightly higher per-token costs and slightly less optimal capability matching, but doesn't break the procurement strategy entirely.
Whether Meta ships Llama 5 in the next two quarters is one of the most consequential strategic questions for the open-weight ecosystem. The longer the silence, the more the answer settles into "the open-weight frontier is permanently a non-US story" — which has consequences for US AI policy and federal procurement that the regulatory regime hasn't fully processed.
Presenc AI — Open-Source LLM Landscape 2026 15+ Model Families → · Hugging Face — Best Open-Source LLM Models 2026 → · O-Mega — Top 10 Open Source LLMs 2026 DeepSeek Revolution →