Topic
38 articles tagged Open Weights. For the wider topic, see Local & Private AI.
Alibaba released Qwen3.8-27B in August 2026 with native vision, a 262K context window, and an Apache 2.0 license. What that actually changes if your work is text.
Read article →Most new open-weights models now read images as well as text. Whether that matters for writing, and why the model that rewrites your emails can stay small.
Read article →Z.ai shipped GLM-5.3-Flash under plain MIT, then released the larger GLM-5.3 two days later under a different, revenue-gated license.
Read article →Alibaba's August 2026 preview of its next model architecture ships open weights, but not under MIT or Apache. What the license actually restricts.
Read article →Meta released a 30B Apache 2.0 model on 10 August 2026 that runs on one GPU. Which machines that means, and whether it helps if your job is writing emails.
Read article →A bigger model is not automatically the better one for rewriting. A fifteen minute test using your own writing that settles it properly.
Read article →Local models do not auto-update, which is a feature. When a new version is worth the download, how to test it against your own writing, and when to skip it.
Read article →Open weights is not open source, not a promise of privacy, and not a licence to do anything. What the term actually covers, and the four things people wrongly assume.
Read article →The open model supply has concentrated sharply this year. A dated scoreboard of who is publishing, under which licenses, and what it means for procurement.
Read article →The project that made local models practical crossed a milestone in 2026. What it actually does, why it matters for privacy, and what it means that it is a dependency.
Read article →Mistral put a new open-weight model into early access without disclosing specifications. Why the European supply question matters separately from benchmarks.
Read article →DeepSeek shipped an MIT-licensed 284B mixture-of-experts model on 31 July 2026. What the license actually gives you, and why cheap hosting is the real story.
Read article →What GGUF files are, how to read the cryptic names, what Q4_K_M actually means, and the three things to check before you download several gigabytes.
Read article →Public benchmarks measure coding and mathematics. Here is a twenty-minute test that measures whether a model will actually help with your email.
Read article →Three paths from nothing to a working local model, what each one costs you, and the five settings that determine whether it feels fast or unusable.
Read article →Four-bit quantization is now the default for local models. What it costs in quality, where the floor is, and why the answer depends entirely on your task.
Read article →Nearly every large model in 2026 is a mixture of experts. What that means, why the two parameter counts matter differently, and the planning mistake it causes.
Read article →Small models handle some languages far better than others. How to test coverage on your own text, and which families are worth trying first.
Read article →A practical shortlist of open-weights models that actually help with professional writing, sorted by what hardware you have rather than by benchmark score.
Read article →Open weights at frontier scale sound like they put frontier AI on your desk. The arithmetic says otherwise. What the memory math actually allows, and what you should run instead.
Read article →DeepSeek V4 competes on inference efficiency rather than raw benchmark scores. Why that is the more interesting strategy, and what it means for anyone paying for AI by the token.
Read article →Gemma 4 moved to Apache 2.0 and ships sizes built for laptops rather than datacenters. What it is good at, where it fits against Qwen, and why on-device is the interesting category.
Read article →Zhipu's GLM line built its reputation on function calling and structured output rather than benchmark scores. Why that reliability is harder to achieve than raw capability, and where it matters.
Read article →OpenAI ships gpt-oss under Apache 2.0 in 120B and 20B sizes. Where they fit, why a reasoning-oriented model is a mixed blessing for rewriting, and how they compare to the small Qwen and Gemma models.
Read article →A decision procedure that starts with your actual RAM instead of a leaderboard. Which size, which variant, which quantization, and how to know when you have picked wrong.
Read article →Model cards and system cards are the closest thing to a datasheet AI has. What to look for, what the omissions tell you, and why this is becoming a compliance document.
Read article →Thinking Machines Lab shipped its first open-weights model in July 2026. Why a US entry matters in a field that had tilted heavily toward Chinese labs.
Read article →Moonshot AI released Kimi K3 as a 2.8-trillion-parameter open-weights model in July 2026. What that actually means, who can run it, and why it changes the field even if you never touch it.
Read article →MiniMax M3 handles a million-token context at a fraction of standard transformer compute. How sparse attention works in plain terms, and why long context is cheaper but not better.
Read article →For European organizations, where a model runs is a compliance question. How Mistral's open lineup fits, and why sovereignty is solved by architecture more often than by geography.
Read article →NVIDIA keeps shipping compressed open-weight models derived from larger ones. How pruning and distillation work, and why compressed models are the ones that reach your laptop.
Read article →SWE-bench, MMLU-Pro, and AIME scores predict almost nothing about whether a model will rewrite your email well. What to measure instead.
Read article →Apache 2.0, MIT, Llama community licenses, and research-only terms give you very different rights. A plain-English guide to what you can legally do with a downloaded model.
Read article →Cloud inference prices fell roughly 80 percent between 2025 and 2026. Here is what that does to the local-versus-cloud calculation, and why cost is now the wrong reason to choose either one.
Read article →Open weights, open source, and open license are three different claims. What each one gives you, what it withholds, and which one matters if you want AI that runs on your own machine.
Read article →Alibaba ships Qwen models faster than anyone can track. A guide to which sizes matter, what the naming means, and why the small ones are the important ones.
Read article →Hunyuan 3.0 shipped open weights in July 2026 with three selectable inference modes. Why letting the user choose how hard the model thinks is the most practical feature of 2026.
Read article →Open weights turn a data-handling promise into an architectural fact. Why that distinction is the whole argument for anyone writing client, patient, or contract text.
Read article →