1) Training: a lab (Meta, Mistral, DeepSeek, Moonshot, TML) conducts a frontier training run โ typically weeks/months on clusters of 1000+ accelerators, costing $10M-$100M+ (Llama 3 405B: ~30M GPU-hours on H100). 2) Post-training: fine-tuning on instruction data, RLHF/RLAIF safety alignment, red-teaming iterations. 3) License decision: choice of license (Apache 2.0, MIT, custom). 4) Publication of assets: (a) weights uploaded to Hugging Face (safetensors format, GGUF for llama.cpp) and mirrored on other hosts; (b) model card with description, benchmark scores, safety evals, use-case guidance; (c) inference code (transformers, vLLM, SGLang integrations); (d) optionally: quantized versions (int8, int4, MXFP4) for smaller GPUs. 5) Ecosystem uptake: community fine-tunes (Alpaca-style instruction tuning, LoRA adapters), quantizations (Q4_K_M, IQ2_XXS), inference optimizations (vLLM, TensorRT-LLM), evaluations on benchmarks (Open LLM Leaderboard). 6) Downstream integration: firms and individual users incorporate the model into their own products (RAG, agentic, chat interfaces, sandbox research). 7) Long tail: models are available literally FOREVER โ weights once uploaded cannot be "undone", hence the regulatory arguments against open-weight at high capability tiers.
Closed/proprietary AI (access only via API) creates three fundamental problems: (1) vendor lock-in โ client dependent on API provider for price, availability, usage policy, deprecation. (2) independent safety research โ external researchers cannot probe weights or intermediate activations, limiting expertise. (3) technological sovereignty โ countries and firms don't want to depend on another country/firm's infrastructure for critical use cases (defense, medicine, finance, intelligence). Open-weight solves all three: weights can be downloaded once and run forever on own infrastructure, researchers have full access to internals, countries can build their own stacks. Additionally: (4) cost โ local inference $0.10-$1.00 per 1M tokens vs $10-$50 for frontier APIs; (5) privacy โ data never leaves client infrastructure; (6) customization โ fine-tuning on own domains/use cases (not possible with API models).
Apache 2.0 / MIT: unrestricted commercial use, critical for enterprise adoption. Llama Community: OK with 700M MAU restriction. OpenRAIL-M / custom: use-case restrictions may exclude legitimate use cases. License choice dramatically affects commercial adoption.
OSAID (OSI, October 2024) requires weights + code + training data information. Most open-weight models provide only weights + inference code (no training data). AllenAI OLMo, MOSS are rare exceptions meeting OSAID. Trade-off: full transparency vs. competitive advantage for the creator.
Immediate release with day-1 access (Kimi K3, Inkling). Staged release: smaller versions first, then larger (Llama series). Delayed release: 3-6 months after internal safety evaluation. Depends on capability tier and risk assessment.
Provider may release only fp16 (community will generate quants) or pre-quantized variants (int8, int4, MXFP4). Kimi K3 releases MXFP4 natively. Greater permissiveness increases adoption.
Open-weight models may have safety layers removed by fine-tuning ("abliterated" models on Hugging Face). Provider can either (a) accept, (b) prohibit via license (ineffective enforcement), (c) build interpretability-based robustness. Trade-off freedom vs safety.
Minimal (name, license) vs comprehensive (architecture, data, benchmarks, safety, limitations, use cases). EU AI Act requires comprehensive for GPAI. Best practice: full model card + system card for safety-critical models.
Model execution (inference) depends on its architecture (dense transformer, MoE), not on open-weight vs closed status. Many open-weight models 2024-2026 are MoE (DeepSeek V3, Mixtral, Kimi K3) with sparse activation โ open-weight is distribution metadata, not architecture.
The distribution itself is parallel (mirror hosts, CDN). Local inference scales to the client's available GPUs. Community fine-tuning and quantization are embarrassingly parallel workloads. Open-weight enables massive parallel ecosystem development โ thousands of independent deployments in parallel.
