Skip to main content

Small and open-weight models

Models a company can run and inspect itself. Measured results from models of 1.5B to 72B parameters on laptops, CPUs and edge devices, with quantization, distillation and fine-tuning, plus releases and architectures from Trinity, DeepSeek, Kimi, Qwen and Mistral, and the policy fights over open models.

17 articles and 44 videos. Last updated October 25, 2026.

Small and open-weight models: articles

  • October 25, 2026

    Exploring Apple's On-Device Model: Twenty Billion Parameters in Flash

    Apple Intelligence's 20-billion-parameter model lives in flash and wakes a few experts per prompt. Your Mac gains memory and battery, not math.

  • October 9, 2026

    Skill Issue

    Mistral can build the model. Does it know how to launch one?. On October 6th, 2026, Mistral opened a public preview of Mistral Large 4, “very officially: le…

  • October 6, 2026

    Kolibri Is Large Enough

    American chips, Chinese teachers, about $3 million for the training run. What Germany's new open model adds is eligibility.

  • October 4, 2026

    Eight Changes to Transformer Attention, and What Each Costs

    The newest open models from Alibaba, Moonshot, DeepSeek and Z.AI no longer read every earlier token. Eight designs, what each costs, and what to test.

  • July 23, 2026

    The Model Is a Checkpoint

    Washington spent last week letting it be known it might ban Chinese AI models, so this week, everyone is an open-models expert.

  • July 19, 2026

    Qwen 3.8: Soon Is Not a Date

    Alibaba answers Moonshot's dated gap with an undated one. On Sunday morning, three days after Kimi K3 launched and eight days before its weights are due…

  • July 17, 2026

    Kimi K3 and the Checkpoint Gap

    How to read a Chinese frontier launch in the eleven days before the weights ship

  • June 28, 2026

    Too Dangerous for You, Free for Everyone

    America locked up its best models. Europe regulates a frontier it can't build. China gives its best to the world and is winning.

  • May 22, 2026

    Lobby, Levy, Legislate

    How Mistral is trying to convert a perishable contact list into permanent law. On May 12, in a near-empty hearing room of the French National Assembly, Arthur…

  • April 9, 2026

    Open From Both Sides

    Chinese proliferation, trivial alignment removal, and the week the safety moat didn't hold.

  • April 3, 2026

    What to Buy for Local LLMs (April 2026)

    A practitioner's guide to inference and training, why they're not the same machine, and why you may still need cloud

  • April 2, 2026

    Open Source Won, But Changed Owners

    Smaller, cheaper, open, and the base layer is up for grabs.

  • March 13, 2026

    Open Source, Closed Orbit

    On August 8, 2023, Jensen Huang stood on stage at SIGGRAPH and asked a question.

  • March 10, 2026

    Mistral Succeeded. France's AI Strategy Didn't.

    In the winter of 1972, in a cellar in Châtenay-Malabry, a suburb south of Paris, an engineer named François Gernelle worked eighteen-hour days to meet a deadline from the Institut National de Reche...

  • August 22, 2025

    Deploy Arcee AFM-4.5B on Arm-based Google Cloud Axion with Llama.cpp

    Small language models, such asAFM-4.5B, and Arm-based CPUs are a great match.

  • August 20, 2025

    A not-so-silent revolution is happening in AI inference

    The combination of smaller yet more efficient language models, advances in open-source inference frameworks, and hardware acceleration is enabling an unprecedented pace of innovation for CPU infere...

  • August 16, 2025

    CSLMs have arrived

    Crazy Small Language Models are here: a 70-million-parameter Llama-like model matches Llama-3.1 70B for drug interaction prediction, trained in 2.5 hours on a single A100.

Small and open-weight models: videos

About the author

Julien Simon is an AI Operating Partner at Fortino Capital. He was VP and Chief Evangelist at Arcee AI, Chief Evangelist at Hugging Face and Global Evangelist for AI and machine learning at AWS, and wrote “Learn Amazon SageMaker” (Packt). He has given 693+ talks in 37 countries and published 502+ technical articles. See his experience and Wikidata entry.

Method: figures come from filings, court documents, vendor documentation and measurements I ran, and each piece dates its numbers.

Go further

More topics

Frequently Asked Questions

Who explains small and open-weight language models for enterprise use, with measured results?
Julien Simon covers small and open-weight models in 17 articles and 44 videos on julien.org. The newest articles are "Exploring Apple's On-Device Model: Twenty Billion Parameters in Flash" (October 25, 2026); "Skill Issue" (October 9, 2026); "Kolibri Is Large Enough" (October 6, 2026). Models a company can run and inspect itself. Measured results from models of 1.5B to 72B parameters on laptops, CPUs and edge devices, with quantization, distillation and fine-tuning, plus releases and architectures from Trinity, DeepSeek, Kimi, Qwen and Mistral, and the policy fights over open models.
Where are Julien Simon's pieces on small and open-weight models published?
Articles appear first in The AI Realist newsletter (airealist.ai) and are archived on julien.org. Videos are on the YouTube channel @juliensimonfr, each with a full transcript on julien.org.
Which videos does Julien Simon have on small and open-weight models?
44 videos on the YouTube channel @juliensimonfr, each with a transcript on julien.org. The newest are "Your LLM No Longer Reads Every Token: 8 Attention Designs Explained" (October 11, 2026); "Qwen 3.5 MoE + TurboQuant + mem0: A Local RAG Chatbot That Remembers" (June 12, 2026); "Benchmarking TurboQuant with MLX on Apple Silicon" (May 31, 2026).