Model Families

Meta's open-weight LLM family — the dominant baseline for self-hosted text generation. Llama 3.3 70B is the canonical 70B-class chat model in 2026; Llama 3.1 8B remains the most-deployed sub-13B production model.

Alibaba's flagship open-weight family with permissive licensing across most variants. Qwen 3 235B-A22B is the leading open-weight MoE for production reasoning; Qwen 3 32B dense is the strongest 32B-class chat model.

DeepSeek's frontier reasoning + code MoE family. DeepSeek V3 (671B MoE) and V4 are the leading open-weight reasoning models in 2026; DeepSeek Coder V3 is the canonical open-weight code model. All ship under DeepSeek's permissive commercial-friendly license.

Mistral's mixed open + closed family. Mistral 7B + Mixtral 8x22B are the open-weight standards; Mistral Large + Codestral are commercial. Codestral Mamba 7B introduced state-space models to production code workflows.

Google's open-weight derivative of Gemini research. Gemma 2 + Gemma 3 cover sub-30B chat; CodeGemma adds code-specialized variants. Tight integration with Vertex AI / Google Cloud / Android Studio.

Microsoft's small-but-strong family — Phi-4, Phi-3.5 lineage. Trained on synthetic data for reasoning. The canonical 'punching above weight class' family — Phi-4 14B competes with 70B-class models on reasoning benchmarks.

Cohere's RAG-tuned open-weight family with first-class document citation + multilingual coverage. Non-commercial license is the practical limit vs Llama / Qwen for production use.

Cohere For AI's multilingual research family. Aya 23 + Aya Expanse cover 23+ languages with explicit balance — the strongest open-weight multilingual chat models for underserved languages (Arabic, Korean, Hebrew, Vietnamese).

01.AI's open-weight family. Yi 1.5 34B was a strong dense model; Yi-Lightning shifted to MoE. Permissive Apache 2.0 license; strong Chinese-English capability.

NVIDIA's reasoning-tuned family. Nemotron-3, Nemotron-4 lineage. NVIDIA-aligned tooling integration (NeMo, TensorRT-LLM); strong on agentic + reasoning workloads.

Databricks' MoE family — DBRX Base + DBRX Instruct. 132B total / 36B active MoE. Surpassed by 2026 MoE leaders but remains relevant for Databricks platform integration.

Baidu's family — ERNIE 4.0, ERNIE Turbo. Strong Chinese-language capability; mixed licensing. Lower deployment in Western open-weight ecosystems; relevant for Chinese-market deployments.

Chinese-market focused open-weight family. Baichuan 4 + earlier versions; strong Chinese capability with permissive licensing.

OpenAI's closed-weight family — GPT-4, GPT-4o, GPT-5 series. Reference baseline for capability comparisons; not self-hostable. RunLocalAI covers GPT family for benchmark context only.

Anthropic's closed-weight family — Claude 3 Haiku/Sonnet/Opus, Claude 3.5/3.7 Sonnet. Reference baseline for reasoning + tool-use comparisons; not self-hostable. RunLocalAI covers Claude for benchmark context.

Google's closed-weight family — Gemini Pro, Gemini Ultra, Gemini Flash. Open-weight derivatives ship as [Gemma](/families/gemma). Reference baseline for capability comparisons.

Alibaba's multimodal vision-language family. Qwen2.5-VL, Qwen2-VL — the canonical open-weight VLMs for OCR + document understanding + chart reading + UI analysis. Qwen2.5-VL beats specialized OCR engines on complex layouts.

OpenGVLab's open VLM family. InternVL 2.5 series spans 1B to 78B. Strong multilingual vision capability; dense alternative to Qwen-VL.

The pioneering open-weight VLM family. LLaVA-1.5, LLaVA-Next, LLaVA-OneVision. Established the VLM training recipe that Qwen-VL + InternVL refined.

Black Forest Labs' rectified-flow image-gen family. Flux Dev + Flux Schnell are the open-weight leaders for text-to-image in 2026; text rendering quality is meaningfully better than SDXL/SD3.5.

The pioneering open-weight image-gen family. SDXL remains widely deployed; SD 3.5 Large is the architectural successor. Massive finetune ecosystem (Pony, Illustrious, NoobAI, dozens of community models).

Alibaba's open-weight video generation family. Wan 2.1 covers text-to-video and image-to-video at frontier-tier open-weight quality.

Tencent's HunyuanVideo + Hunyuan3D family. HunyuanVideo is a 13B parameter open-weight video model; Hunyuan3D-2 is a leading text-to-3D + image-to-3D system.

OpenAI's open-weight speech-to-text family. Whisper Large v3 + Whisper Turbo are the canonical open-weight STT models. faster-whisper + WhisperX deliver production-grade speed via CTranslate2 backends.

BAAI's open-weight embedding family. BGE-M3 is the canonical multilingual embedding model in 2026 (100+ languages, 8K context); BGE Reranker V2 M3 is the canonical companion cross-encoder reranker. The default open-weight RAG retrieval stack.

The BigCode community's open-weight code family. StarCoder 2 + earlier StarCoder. Permissive license + transparent training data — the 'reproducible code model' canonical pick.