Jul 1, 2026 / AI Infrastructure
The three credible local ASR options on Apple Silicon in mid 2026 are Whisper Large v3 Turbo, NVIDIA Parakeet TDT, and Distil Whisper. We ran all three on a 128GB M5 Max under mlx-whisper across 40 hours of real audio, then measured word error rate, real time factor, streaming latency, peak memory, and the cost per audio hour served. The model with the lowest WER is not the model that wins the cost per audio hour line.
Jun 30, 2026 / AI Tooling
All three frontier models now advertise 1M token context windows, but advertised context length and usable retrieval at depth are different metrics. We ran a 12,000 query needle in a haystack benchmark across Fable 5, Claude Opus 4.8, and Gemini 3.5 Pro at context depths from 128K to 1M tokens, with single needle and multi needle variants, and measured recall, latency, and cost per successful recall. The model that wins the marketing line does not win the cost per recall line.
Jun 30, 2026 / AI Infrastructure
The KV cache is the silent driver of context length on Apple Silicon. A Llama 3.3 70B model holding 512K context across a single session needs roughly 80GB just for the KV cache at FP16, which blows past the working budget on a 128GB M5 Max once the model weights are loaded. Offloading inactive sessions' KV cache to NVMe SSD instead of evicting them is the pattern that turns a single Mac Studio into a many session long context server, but the design only works inside a narrow operating envelope.
Jun 30, 2026 / Headless Commerce
Migrating off Shopify Plus to an open source headless backend is not a single project. It is a sequence of nested cutovers, each with its own failure modes and its own rollback story. The plugin architecture of the target platform decides how much of your custom Shopify Plus logic ports over cleanly, the data migration order decides whether the cutover window is hours or weeks, and the dual write window decides whether you can roll back if something goes wrong. The hard parts are not on the brochures.
Jun 29, 2026 / AI Tooling
The benchmarks that decide which frontier model goes into a strict tool calling workload are not the headline coding evals. They are the schema adherence rate on deeply nested arguments, the malformed call rate at scale, the latency on a single tool call, and the per call cost. We measured all four across Claude Opus 4.8, GPT 5.5, and Gemini 3.5 Pro on 1,500 production tool calls, and the ranking is not the one the marketing pages predict.
Jun 29, 2026 / AI Infrastructure
Tool calling is the workload that exposes whether a local model is actually agent ready. We measured strict JSON schema adherence, malformed argument rate, and recovery behavior on Qwen3 32B Instruct and Llama 3.3 70B Instruct under mlx-lm on a 128GB M5 Max across 1,200 synthetic tool calls drawn from real agent traces, and the gap between the two models is larger than the published evals suggest.
Jun 29, 2026 / AI Tooling
The cost per resolved support ticket is the metric that decides which frontier model goes into a triage agent loop, not the per token list price. We ran 6,000 tickets from a real Shopify Plus support queue through three pipelines (Opus 4.8 routing only, GPT 5.5 end to end, Haiku 4.5 with Opus 4.8 escalation) and measured auto resolution rate, escalation precision, average latency, and total cost per ticket. The cheapest list price is not the cheapest pipeline.
Jun 29, 2026 / AI Infrastructure
A 512GB M5 Ultra Mac Studio is the first Apple Silicon machine that can host a 235B parameter MoE model in unified memory without quantizing into uselessness. We measured Qwen3 235B MoE (22B active) and Llama 3.3 70B Dense head to head under mlx-lm on the same machine across decode throughput, prefill latency, peak memory, sustained throughput across a one hour run, and the cost per million tokens served. The sparse vs dense tradeoff does not land where the parameter count alone predicts.
Jun 29, 2026 / Headless Commerce
A Shopify Hydrogen storefront serving a brand that publishes editorial content daily lives or dies on the CMS choice. Storyblok and Sanity are the two credible options once you rule out the visual-builder products and the enterprise stacks; the choice between them comes down to visual editing fidelity, preview latency, and the content ops cost that compounds across hundreds of pages per quarter.
Jun 29, 2026 / Headless Commerce
Webhooks are how a headless commerce backend learns about the events that move money, change inventory, and update the customer record. The three platforms that drive most enterprise stacks (Shopify, Stripe, NetSuite) have radically different webhook reliability stories, and the difference shows up at the worst possible moment: the day the pipeline drops an order or double-charges a customer. The retry policies, idempotency surfaces, and ordering guarantees are not interchangeable.
Jun 28, 2026 / AI Tooling
Repo migration is the hardest workload for an agentic CLI because the task lives across hundreds of files, the success criterion is a green test suite plus a clean diff, and the model has to hold context across a multi hour session. Claude Code 2, OpenAI Codex CLI, and Aider all compete for the seat, and the completion rate, cost per merged PR, and diff quality numbers move in different directions across the three.
Jun 28, 2026 / AI Infrastructure
MLX 0.21 shipped an FP8 path that finally gives Apple Silicon a quantization tier between Q6 and full precision without the rounding artifacts that Q4 imposes on long sequences. We measured tokens per second, peak unified memory, and quality regression across three eval suites on a 128GB M5 Max, and the choice between FP8, Q6, and Q4 in a production serving path is not the obvious one.