Sep 18, 2026 / AI Infrastructure
Alibaba released the weights of Qwen3.8-Max on August 12, 2026, as Qwen3.8-2.4T-A95B with FP8 checkpoints on Hugging Face, which means you can legally run a model that is at parity with frontier systems on your own hardware. Then the arithmetic starts. At 2.4 trillion parameters, the FP8 weights alone occupy roughly 2.4 terabytes of memory before you allocate a single token of KV cache, and the mixture of experts design that keeps compute cheap does nothing to shrink that footprint, because all of the weights have to be resident somewhere even though only 95 billion activate per token. This post does the honest hardware math, covers the quantization and expert offloading options that change it, and works out the break-even against the hosted API at 2 dollars and 6 dollars per million tokens.
Sep 18, 2026 / Commerce Engineering
Shopify Functions can be written in any language that compiles to WebAssembly, and the two most common choices, Rust and JavaScript, are not close on performance. JavaScript Functions run about three times slower than Rust, because the JavaScript runtime itself has to be bundled into the Wasm module, while Rust and AssemblyScript compile more or less directly to Wasm instructions. That gap would be academic if Functions had generous limits, but they do not: Shopify enforces an 11 million instruction budget per invocation, and blowing past it means Shopify rejects your output and falls back to default behavior. For simple discount rules the language does not matter. For complex logic over large carts, it decides whether your Function runs at all. This is the trade-off, with real numbers and a concrete pattern for staying inside the budget.
Sep 18, 2026 / Commerce Engineering
Shopify disabled creating and editing Scripts on April 15, 2026, and Scripts stopped executing entirely on June 30, 2026. Shopify has confirmed the deadline will not move again. If your Plus store still has discount logic, shipping rules, or payment method gating that lived in a Ruby Script, that logic is not running anymore, and your checkout is quietly falling back to Shopify defaults. The replacement, Shopify Functions, is not a port of the same code. Functions are compiled WebAssembly modules that run inside Shopify's own infrastructure, and the migration is an architecture change, not a copy and paste. This playbook covers what actually broke, how Functions differ, and the order to rebuild in so the highest revenue logic comes back first.
Sep 17, 2026 / AI Models
On September 11, 2026, Shanghai AI Laboratory released Atria Dawn Preview, a 744 billion parameter agentic mixture of experts model with a 256K context window, shipped as open weights under the MIT license with no hosted per token price at launch. The framing is a research agent built for long horizon analysis, code, and tool use, and the paper claims it is competitive with frontier agents across sixteen benchmarks and posts the highest reported score on five of them, including 96.0 percent on the lab's own DeepSearchQA and 59.6 percent on SWE-bench Pro. Those are attention grabbing numbers and they all carry the same asterisk: the lab ran or commissioned every eval, and none has been independently reproduced as of launch. A 744B open weight model with genuinely strong self reported agentic scores is worth taking seriously and worth being skeptical about at the same time, because the open weights are a real gift and the benchmark table is a marketing document until someone else runs it. This post separates what is verifiable from what is a claim, and works out where a model this size actually fits given that running it is a serious infrastructure commitment.
Sep 17, 2026 / AI Models
Anthropic made Claude Fable 5.1 generally available on September 1, 2026, and within two weeks the independent BenchLM coding rankings had it on top of SWE-bench Verified at 83.9, ahead of Claude Fable 5 at 77 and Claude Opus 5 at 75.7. That top line is the part everyone quotes, and it is also the part that matters least for a build decision, because a coding agent that runs thousands of times a day is judged on cost per solved issue, not on a leaderboard row. A model that passes seven points higher but burns twice the tokens on every attempt can be the more expensive way to close the same ticket. We ran both models through the same fixed harness on SWE-bench Verified style tasks, held the scaffold and retry policy constant, counted tokens per solved issue rather than tokens per attempt, and turned the pass rate and the token spend into a single number you can actually budget against. The pass rate gap is real and it favors Fable 5.1. The cost picture is more interesting, because Opus 5 and Fable 5.1 fail in different places and spend tokens differently on the way to a fix, and that is where the decision actually lives.
Sep 17, 2026 / AI Tooling
The question people keep asking, Cursor or Claude Code or Codex, has quietly stopped being the right question. Through 2026 these three did not converge into a single winner and they did not stay interchangeable either. They settled into different jobs. The pattern that developers actually run, and that the trade press has started calling a composable coding stack, splits the work into layers: an editor where a human stays in the loop and reviews, a terminal native agent that runs long autonomous jobs across a whole repository, and a cloud delegated worker you hand a scoped task and walk away from. Cursor, Claude Code, and Codex each landed in one of those layers because of what they are built around, not by accident, and teams that treat them as one stack rather than three competitors ship faster than teams still trying to crown one. This post lays out the three roles, why each tool fits the role it fits, and how to divide a real workday's work across them without paying three subscriptions to do the same thing three times.
Sep 17, 2026 / AI Infrastructure
The Mac Studio with M5 Ultra reaches customers on September 22, 2026, and the configuration worth talking about is the 512GB one, because it crosses a memory line that no consumer machine has crossed before. At full precision, weights alone, a 70B model wants roughly 140GB and a 120B class model wants roughly 240GB, and once you add the KV cache for a working context window the 120B number climbs past what any prior Apple Silicon box could hold. 512GB of unified memory changes that: for the first time you can keep a 120B class model resident in FP16, feed it a long context, and still have headroom, on a machine that sits on a desk and draws wall power instead of a rack. That is a genuinely new capability, not a faster version of an old one, and it is also easy to oversell. FP16 at that size is memory rich and bandwidth bound, so the question is not can it fit, it obviously can, the question is what throughput you actually get and when a quantized model on cheaper hardware is the smarter buy. This post walks the memory math, the prefill and decode behavior, and the cases where 512GB earns its price.
Sep 17, 2026 / Headless Commerce
For most of its life Hydrogen came with a deployment target attached: you built on Hydrogen, you shipped to Oxygen, and the two were sold as one decision. At Editions Spring 2026 Shopify previewed the change that ends that coupling, letting Hydrogen's commerce building blocks run inside Next.js, Astro, SvelteKit, or Nuxt and deploy to Vercel, Cloudflare, or Netlify instead of only Oxygen. Paired with Hydrogen's move onto React Router 7 and its shift from monolithic starter templates to smaller composable recipes, this turns Hydrogen from a framework you adopt wholesale into a set of parts you can drop into a stack you already run. That is a real strategic shift for enterprise headless builds, because the deployment target has always been one of the load bearing decisions in a Shopify Plus architecture, and it was previously made for you. But portable does not mean free, and leaving Oxygen means giving up things Oxygen did quietly on your behalf. This post is about what actually ports, what breaks or costs more off Oxygen, and how to decide whether the freedom is worth it for your build.
Sep 16, 2026 / AI Infrastructure
DeepSeek shipped V4.1 Flash on September 10, 2026, and four days later it did something more consequential than a normal release: it pointed the deepseek-v4-pro endpoint at the new model. From 04:00 UTC on September 14, every request you send to V4 Pro is served by V4.1 Flash, at Flash prices, until a future V4.1 Pro ships. The new model is much smaller, a 552 billion parameter mixture of experts against the old 1.6 trillion parameter model, it ships open weights under MIT, and it is cheaper by a wide margin. It also beats the retired model on agentic coding and tool use while giving up ground on knowledge heavy reasoning without tools. If you have code calling deepseek-v4-pro in production, your model changed under you this week. This post is what actually changed and what to do about it.
Sep 16, 2026 / AI Infrastructure
GLM 5.2 is the model to beat if you want open weights and you write code for a living. Zhipu released it in June 2026 as a 753 billion parameter mixture of experts with a 1 million token context and an MIT license, and it posts the strongest coding numbers of any open model this year: 62.1 on SWE-bench Pro, ahead of GPT-5.5, and 81.0 on Terminal-Bench 2.1, within striking distance of Claude Opus 4.8. It lists at roughly 1.40 dollars per million input tokens and 4.40 per million output, which is a fraction of frontier closed pricing, and the weights are on Hugging Face with no regional limit. As the new cheap hosted tiers arrive in September, the question is not whether GLM 5.2 is good, it clearly is, but where it actually fits against both the closed frontier above it and the cheap agentic models now crowding in below. This post works that out.
Sep 16, 2026 / AI Infrastructure
OpenAI shipped GPT-6 Astra on September 3, 2026, at 10 dollars per million input tokens and 50 per million output, which is 2.5 times the GPT-5.6 Sol rate it sits above. Cached input reads drop to 1 dollar. The benchmark gains are real and, on some tasks, enormous: Terminal-Bench 4.0 jumps from Sol's 37.3 to 57.9, and FrontierMath Tier 4 goes from 83.0 to 97.6. On other tasks the gain is a rounding error: DeepSWE v1.1 moves from 72.7 to 74.1. A 2.5 times price increase for a benchmark that moved 1.4 points is a bad trade, and a 2.5 times increase for a benchmark that moved 20 points can be the best money you spend all quarter. The premium tier is not worth it or not worth it in general, it is worth it per task, and this post does the arithmetic.
Sep 16, 2026 / Engineering For a decade the discovery question for an ecommerce brand was how to rank on Google. In 2026 there is a second surface that behaves nothing like search: AI shopping assistants like ChatGPT and Perplexity that answer a buyer's question with specific products and links, without the buyer ever visiting a results page. Shopify's answer is Shopify Catalog, the channel that exposes your products to those assistants, and since the Winter 26 edition it explicitly covers headless Hydrogen storefronts running on Oxygen, which were previously invisible to this kind of tool. That last part matters, because going headless historically meant giving up the built in surfaces a themed Shopify store got for free. This post is a practical read on how the Catalog channel works, what your product data has to look like to win in it, and how to think about AI shopping as a discovery channel rather than a mystery.