Thinking Machines Lab finally shipped an open model — Inkling, a 975B open-weights MoE. PrismML squeezed a 27B model down to 3.9GB so it runs on an iPhone. Google put real ML inference in the browser. And Europe got a serious open bilingual foundation model.
Mira Murati's lab shipped its first open model — and made it open-weights. Inkling is a Mixture-of-Experts transformer with 975B total / 41B active parameters, native reasoning over text, image, and audio, and a context window up to 1M tokens. Its headline feature is controllable thinking effort — you dial how long it reasons against cost. Not #1 on every benchmark, but positioned as the best open base to customize.
So what: This is the thesis from last week's manifesto turned into a real artifact. A frontier-scale, open-weights, natively-multimodal model built to be fine-tuned is exactly the "customizable weights" bet — and it's the strongest open challenger to closed frontier labs yet.
Vibe by Mistral’s Code Mode launches remote coding agents from a dedicated web surface. Connect to GitHub, manage your projects, and see coding sessions through to a pull request. Sessions run in parallel, persist while your machine is off, and each runs in an isolated sandbox. A new VS Code extension brings the same agent into your IDE, reading, editing, and executing commands beside your files. In the CLI, /teleport moves a live session between your terminal and the cloud, keeping history and approvals intact.
sponsored
PrismML compressed Qwen3.6-27B into low-bit builds with no re-pretraining and no multi-GPU rig. The ternary build (1.71 bits/weight) keeps 94.6% of FP16 quality at 5.9GB; the 1-bit build holds 89.5% at 3.9GB — small enough to run locally on an iPhone 17 Pro, with 262K context and vision input intact.
So what: 27B-class intelligence running entirely on a phone, offline, is a different product category than API calls. If sub-2-bit quantization keeps landing at ~90% quality, on-device becomes the default for privacy-sensitive and latency-critical apps.
LiteRT.js is a JavaScript binding of LiteRT that runs .tflite models inside the browser, accelerated by WebGPU. Because inference stays local, Google cites enhanced privacy, zero server costs, and ultra-low latency. Works in Chrome/Edge 113+ and Safari 17.4+.
So what: This collapses the deployment gap for web ML — no Python backend, no inference server, no per-call cost. Any web dev can now ship a real model as a static asset, which massively widens who can build ML features.
A European consortium released an open bilingual base model built on a hybrid Mamba-Transformer MoE architecture, activating 3.2B of 31.6B parameters. Purpose-built for German and English rather than English-first with a multilingual afterthought.
So what: Sovereign, language-specific open models are becoming a real category. A hybrid-architecture base tuned for a specific language pair signals Europe building foundational infra it controls — not just fine-tuning American weights.
SpaceXAI Open-Sources Grok Build: The Rust Agent Harness Behind Its Coding CLI
SpaceXAI open-sourced Grok Build — the terminal coding agent behind its grok CLI — under Apache 2.0. Crucially, it's the harness, not the model: the Rust source for the full-screen TUI, the agent runtime, and the tool layer (file edits, shell execution). It runs interactively, headlessly in scripts, or via the Agent Client Protocol (ACP), and syncs periodically from the internal SpaceXAI monorepo.
So what: Open-sourcing the harness under a permissive license decouples the agent scaffold from Grok's weights — anyone can now run xAI's exact coding-agent runtime on any model.
Every AI product you use is built on two ideas: tokens (what the model reads) and the context window (how much it can hold). Most explanations of them are wrong in the same way.
This one uses real numbers from a real tokenizer. Ten minutes, no prerequisites.
|
AI Dev Signals · Fundamentals Tokens & context windowsWhy the model can't count letters, and why it re-reads your entire chat every time you press enter. |
|
01 |
What the model actually reads |
A language model never sees your text. Before anything reaches it, a tokenizer chops your sentence into chunks and swaps each chunk for a number. The model only ever sees the numbers.
Here is a real sentence, tokenized with o200k_base — the tokenizer behind the current GPT models. Top row is what you typed. Bottom row is what the model receives.
|
Input — "The cat sat on the hamburger."
6 words → 7 tokens. The · marks a leading space — it belongs to the token. What the model receives is just this: [976, 9059, 10139, 402, 290, 139813, 13] |
Rule of thumb: ~4 characters per token in English, or ~0.75 words per token. Measured on real prose it runs closer to 4.9 characters — treat 4 as a safe overestimate when budgeting.
|
The space changes everything Most token demos quietly cheat by dropping the leading space. Same word, two very different results:
Words get split when they are rare in that position, not when they are long. Mid-sentence, common words are almost always one token. |
||||||||||
|
So why can't it count the r's in "strawberry"? The popular answer is that the model sees str + aw + berry and loses track. Tokenize the actual question and that story falls apart:
Mid-sentence, "·strawberry" is a single token. The model isn't juggling three fragments — it's holding one number that has no letters inside it at all. Asking it to count r's is like asking you to count the strokes in a word you only ever heard spoken. The real answer is worse than the myth, and more interesting. |
|
02 |
What the model actually remembers |
Nothing. That is not a joke and it is the whole lesson.
A model has no memory between turns. It is a pure function: tokens in, tokens out. When you send message #20, the app quietly re-sends all nineteen previous messages plus your new one, and the model reads the entire conversation from scratch, as if for the first time. The context window is simply the maximum size of that re-read.
|
A chat where each turn adds ~300 tokens
Ten turns of 300 tokens is 3,000 tokens of conversation — but you were billed for 16,500. Every turn pays for the whole history again. This is why long chats get slower and more expensive, and why prompt caching exists. |
|||||||||||||||||||||||||||
Simplified: real chats vary per turn, and cached input is discounted heavily. The shape is what matters.
|
The window is not empty when you start Before your first word, the context window already holds the system prompt, any tool and function definitions, retrieved documents, and files you attached. On a tool-heavy agent that overhead can run to tens of thousands of tokens. You are always renting the remainder. |
|
What actually happens when it fills up Here is where almost every explainer gets it wrong. Nothing is "quietly erased." The behavior depends entirely on which layer you are talking to:
So "the AI forgot what I told it" almost always means: a piece of software you never saw decided which parts of your conversation were worth re-sending. If you build on the API, that software is you. |
|
03 |
Where the numbers stand |
The context window arms race is over, and it ended in a tie. Every frontier lab has converged on 1M tokens — roughly 750,000 words, about ten novels. The interesting competition moved to cost and reasoning.
~750,000 words. Max output is the real differentiator: 128K on Opus 4.8 and GPT-5.6, ~64K on Gemini 3.1 Pro. |
~150,000 words — still a full-length book. Enough for the overwhelming majority of production tasks. |
~1,500 words. Its bar is empty because 2,048 tokens is 0.2% of 1M — too thin to draw honestly. Windows grew 488× in six years. |
|
A big window is a ceiling, not a promise Models reliably retrieve facts placed at the start and end of a long context, and degrade in the middle — the "lost in the middle" effect. A 1M window means the request won't error out. It does not mean the model used all of it well. Put what matters at the edges, and benchmark retrieval at the length you actually ship. |
|
Key takeaways
|
|
Token counts computed with o200k_base. Context windows verified against vendor documentation, July 2026. Model specs change monthly — check the source before you quote these. Marktechpost · AI Dev Signals |