AIWiki
Malaysia
Back to all articles
AI Foundationsopen sourcelicensingLLMs

Open Weights

4 min readUpdated October 2026
Open Weights
Type
Model release category / licensing practice
Emerged
2023 (mass adoption, LLaMA and Mistral era)
Key families
DeepSeek, Qwen, Llama, Kimi, GLM, Gemma, Phi, Mistral
Common licences
MIT, Apache 2.0, and vendor-specific custom licences
Key distinction
Weights downloadable ≠ open source (training data and code usually closed)
Related
Open source AI, quantization, self-hosting
Open weights describes machine learning models whose trained parameters (weights) are published so anyone can download, run, fine-tune and often deploy them — without calling the developer's API. The term is deliberately narrower than open source: weights alone do not imply that the training data, training code or inference code are available, and licences vary from MIT and Apache 2.0 to vendor-specific terms with revenue thresholds.[1][2]

History

The category emerged in 2023 when Meta's LLaMA weights leaked and were subsequently released officially, alongside Mistral AI's early Apache 2.0 models — demonstrating that compact open models could rival far larger closed ones. Alibaba's Qwen line, published from Qwen3 onward under Apache 2.0, and DeepSeek's models (MIT-licensed weights from DeepSeek-R1 onward) turned open weights into an industrial-scale ecosystem through 2024–2025.[1] Meta itself softened its language from "open source" to "open weight" with Llama 4, whose community licence carries conditions such as withheld multimodal grants for EU-domiciled entities.[1][3]

By 2026 the capability gap had narrowed dramatically: DeepSeek-V4, GLM-5.3, Kimi K3 and Qwen3.8 were trading blows with closed flagships on public benchmarks while shipping downloadable weights, and OpenAI released its own open-weights line (gpt-oss) after years of being API-only.[2][4] In July 2025 Nvidia's Jensen Huang published an open letter urging Washington not to restrict open-weight models, and by 2026 industry signatures to that position had passed fifty as open models stopped being merely the budget option.[4]

Key Concepts

  • Weights vs. source. Downloadable parameters let you run the model; full open source additionally requires open training data and code. Moonshot AI explicitly labels Kimi K3 "open weight" rather than open source for this reason.[4]
  • Licence gradients. MIT/Apache 2.0 (DeepSeek-R1, Qwen3.x, many Mistral models) allow unrestricted reuse; custom licences (Llama Community License, Kimi K3 License, Qwen3.8-Max Licence) add conditions for very large products or model-as-a-service businesses.[1][3]
  • Self-hosting economics. Open weights can be served on owned or rented GPUs, run privately on-device, fine-tuned, quantised and audited — capabilities closed APIs cannot match.[2]
  • Sovereignty and safety trade-offs. The same downloadability that enables national sovereignty and independent research also enables unfiltered misuse; this tension drives the ongoing US/EU debate over open-weight restrictions.[4]

Applications and Impact

Organisations use open weights to keep data inside their perimeter, avoid per-token costs, customise models for domain language, and guarantee availability independent of any vendor's pricing. The releases have restructured the market: hosted APIs of open models undercut closed incumbents on price, self-hosting became viable at every scale from laptop (27B quantised) to cluster (2.8T-parameter Kimi K3), and governments began treating open-weight capability as strategic infrastructure.[2][4]

>See Also

🇲🇾Malaysian Context

🇲🇾 Relevance to Malaysia: open weights are the backbone of Malaysia's sovereign AI ambitions. YTL AI Labs' ILMU, Malaysia's homegrown sovereign LLM with Malay-language capability, is the clearest example — a national model that Malaysians can inspect, host locally and build on without renting foreign intelligence.[5] For Malaysian businesses, open weights translate into concrete PDPA advantages: personal data can be processed by a model running in a Malaysian data centre or on-premises, with no offshore inference. SMEs gain access to capable models at zero licence cost (Qwen, Gemma, Phi run on modest hardware), while MDEC and NAIO framing — sovereign AI, local capacity, skills transfer — aligns naturally with technology that can be downloaded, studied and fine-tuned locally. Caution remains: "open weights" does not mean licence-free, and Malaysian adopters must still check each checkpoint's terms before commercial deployment.

References

  1. ↑Hidekazu Konishi. Open-Weights LLM Release History and Timeline. https://hidekazu-konishi.com/entry/open_weights_llm_release_history_and_timeline.html
  2. ↑Wavect (2026). Open-Weight LLMs 2026: DeepSeek, Qwen, Kimi, GLM, Llama and Muse. https://wavect.io/blog/open-weight-llm-comparison-2026/
  3. ↑Meta AI. Llama 4 Model Card and Use Policy. https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md
  4. ↑Yotta Labs (2026). Best Open-Source LLMs in 2026: What Actually Runs in Production. https://www.yottalabs.ai/post/best-open-source-llms-2026
  5. ↑YTL AI Labs. ILMU: Malaysia's AI. https://www.ytlailabs.com/
Back to all articles