Intelligence is escaping the data center, one laptop at a time. Issue No. 2 | Week of September 7, 2026
This week: the "Flash" tier of open models had its first real bake-off, Perplexity shipped a hand-tuned Mac inference engine and gave it away, Apple started selling the Mac mini as an always-on agent box, and the price of renting an H100 fell while Anthropic's compute tab passed a quarter-trillion dollars.
The Number: 6 billion. Active parameters per token in Qwen3.8-Flash-Next - the model that just swept the quality rows of the workhorse tier.
The workhorse tier gets a bake-off
The frontier models get the keynotes, but production traffic runs on the Flash tier: high-volume models that are good enough and radically cheaper. Regolo put the three leaders head to head this week - DeepSeek V4 Flash, Alibaba's Qwen3.8-Flash-Next, and Zhipu's GLM-5.3-Flash - using model cards and Artificial Analysis measurements. Result: Qwen's 6-billion-active-parameter model sweeps the language, reasoning, and office-agent benchmarks; GLM-5.3-Flash posts the top independent intelligence score; DeepSeek wins on speed and cost and almost nothing else. The licenses are splitting too: MIT for DeepSeek and GLM, Qwen Community 1.0 for Alibaba's - which limits large-scale commercial use. If you're building on these, read the terms, don't assume them. Regolo's comparison | Qwen3.8 Flash Next deep-dive
Perplexity's gift to every Mac: Lily
Perplexity open-sourced Lily on September 2: a Rust-and-Metal inference engine built for exactly one model (Qwen3.6-35B-A3B) on exactly one chip family (Apple Silicon). The payoff for that narrowness: 1.23x faster prefill and 1.35x faster decode than Apple's own MLX-LM running the same model. Lily is the local half of Hybrid Compute, the split mode in Perplexity's Mac app that keeps work on-device until the cloud is actually needed. The open question it raises for the llama.cpp/Ollama/LM Studio world: is single-model specialization now the fastest way to squeeze tokens out of a MacBook - and does everyone shipping their own engine fragment the toolchain? Tech Insider writeup
Apple wants your Mac mini to never sleep
Apple is now officially pitching the new M6 and M5 Pro Mac minis for "always-on agentic computing" - the use case homelab people were already running. The M6 puts Neural Accelerators in every GPU core, with up to 4.8x the M4's LLM prompt-processing in Apple's own LM Studio testing; the M5 Pro goes to 64GB of unified memory, and Apple is explicitly promoting Thunderbolt 5 clustering of multiple minis to run larger models. Meanwhile Meta's new Muse Glimmer, a 30B model designed for local agents, fits in 24-32GB of memory - a typical well-specced desktop. The device story and the model story are converging on the same box. AsiaOne
The money: prices down, bets up
Together cut dedicated H100 instances from $5.49 to $3.99 an hour for September, and one analysis this week pegged the hyperscaler premium over neoclouds for the same H100 at 89%. Compute at the edge of the market keeps deflating. At the center, the opposite: Anthropic's 2026 compute commitments - AWS ($100B over ten years), Fluidstack, Nscale, SpaceX, Lambda - have pushed one pre-IPO company's pledged cloud spend past $275 billion. Both things are true at once: unit costs are falling and the centralized buildout keeps getting bigger. The gap between those two curves is where the edge lives. PulseAugur on Together's cut | Value Add Pulse on the $275B
Distributed Intelligence is sent and paid for by Smarty Pants, a distributed AI inference network that runs open-weight models on idle Apple Silicon and pays Macs for the intelligence they produce.