A 26-billion-parameter model ran on an 8 GB MacBook Air with a reported memory footprint of roughly 2 GB.
Then a thousand-token prompt took 37 seconds before the first word appeared.
Saving memory, adding compute, and building custom chips can all help. But each approach moves the bottleneck somewhere else.
2 GB of Memory Came With a 37-Second Wait
Turbo Field Fair runs a model with 14 GB of weights by taking advantage of Gemma 4’s structure. Each layer has 128 experts, but the router selects only eight for each token, while one shared expert always runs.
The runtime keeps the essential parts in memory. The core uses 1.35 GB. The other weights remain across 30 SSD files, fetched as the router selects them.
That makes the model fit. It also puts SSD access directly into token generation.
For one diagnostic token:
- SSD expert reads: 83 milliseconds
- Graphics command queue: 55 milliseconds
- Output layer: 14 milliseconds
- Total: 162 milliseconds
Caching helped. About 41% of the experts selected for one token were selected again for the next. Caching 16 experts per layer cut repeated expert reads from 166 to 88 milliseconds per token. A different eviction policy saved another 8 milliseconds.
Prediction did not help. An attempt to predict future expert selections reached only 7% accuracy, so the runtime did not use it.
Rewriting the expert math for the GPU worked much better:
239 milliseconds → 60 milliseconds
The repository records 103 experiments, including failures.
On the MacBook Air, Turbo Field Fair generates between 5 and six tokens per second. But the thousand-token prompt still takes 37 seconds before generation begins.
Fitting the Model Was Only Part of the Cost
On the same 24 GB M5 Pro and model:
- Turbo Field Fair: 31 to 35 tokens per second
- MLX: 76 to 82 tokens per second
The roughly 2 GB figure does not capture the full system cost. macOS can use idle memory as a disk cache without charging it to the process. Under memory pressure, the author’s M5 dropped from 35 to 27 tokens per second.
The source recommends MLX when enough memory is available. Turbo Field Fair is another option when an 8 or 16 GB Mac cannot fit the model normally.
A Custom Chip Posted Strong but Limited Results
OpenAI reported that Jalapeño, its Broadcom-built inference chip, delivered up to 1.9x more AI work per watt and up to 3.6x lower latency than Nvidia GB200 and GB300 in internal tests. Reported power use was around 700 watts.
The chip is intended for inference, not training. OpenAI said its advantage increased on larger workloads, including Moonshot’s Kimi model, and that it performed well on unreleased OpenAI models.
These were first-party results. The comparison used GB300, not Nvidia’s newer Vera Rubin generation.
OpenAI plans a small-volume deployment this year and a significant production increase in 2027. Its second-generation chip is approaching tape-out, and work on a third generation has begun.
Apple Is Targeting Two Different Local-AI Machines
The $899 M6 Mac mini has a 12-core CPU, 12-core GPU, up to 32GB of unified memory, and 170GB/s of bandwidth. Apple calls it an “always-on agentic device.”
Apple claims up to 4x faster AI performance than M4 and up to 4.8x faster LLM processing in LM Studio. Another brief reports up to 13.5× faster LLM prompt processing than M1.
The base machine has 16GB of memory and 256GB of storage. Those limits matter for local models.
At the other end, the M5 Ultra Mac Studio has a 36-core CPU, 80-core GPU, up to 512GB of unified memory, and 1.2TB/s of bandwidth.
Apple claims up to 9.8× faster LLM prompt processing than M1 Ultra and 4× faster than M3 Ultra. One source described a $9,499 configuration with 256GB of unified memory as capable of running DeepSeek V4 Flash locally.
One Memory Bus Beat Four Connected Boxes
A comparison placed a 512GB M5 Ultra Mac Studio against four DGX Spark boxes and four AMD Ryzen AI Halo boxes. The systems cost a relatively similar $15-20k.
Four DGX Sparks connected over 200GbE combined compute fastest. But the model was split across the machines, and every token passed through them sequentially. Four boxes still generated at the speed of one.
The Mac kept its 512GB on one bus. The source reported approximately 4x the tokens per second of the Spark or Halo systems, although the Mac felt slower on long prompts.
The Halo systems used 10GbE, which the source said was too slow to share work properly.
The Mac was silent and used less power than a gaming PC. Four Sparks ran hot and nearly maxed out a wall circuit. Four Halos used about half the Sparks’ power but remained somewhat hot.
Fitting a model into memory is not enough. The system must also move the model’s data quickly enough.
Faster Agents Brought More Parts and Visible Failures
The daily social digest reported that Codex CLI 0.149.1 added thread-source classification, image budget enforcement, and memory consolidation tagging. It also reported roughly 25× faster startup after Charlie Marsh reworked the lifecycle.
Collections now package agents, skills, commands, security tests, workflows, and training material. One prompting approach has the agent build a scoring skill first, then improve its work against that score.
The failures are visible too. The digest reported a Codex Desktop bug with GPT-5.6 Sol that caused 400 errors after tool calls, a Claude Code launch segfault, and friction across Cursor Cloud Agents.
Claude Moved Review Work Inside Word
Claude for Word works inside Microsoft Word on the web, Windows, and Mac.
It can read a complete document, connect reviewer comments to their locations, and propose edits through track changes. It asks permission before changing the document’s editing mode.
It can also check claims against Box files. In the example, Claude compared performance and speed claims with the Rockpool 3.0 Early Access Results document, matched claims by meaning, and presented corrections before applying them.
The example preserved a customer quote and availability dates, repaired formatting, and reduced the draft below 1,000 words.
Big Headlines Still Needed Evidence
Several large claims could not yet be evaluated.
One claimed 1,000 tokens per second on a single RTX 3090, but the supplied excerpt did not include the promised configuration or benchmark conditions.
Another described a secret free coding model as beating Claude and GPT while also saying there was no explanation and nobody had claimed it.
A post about Boris prompting 1000+ agents overnight mentioned “graph engineer,” but provided no technical account of the system.
A video said Nvidia agreed to pay $6 billion to rent “the machine inside” a company, based on a private letter sent to a few dozen investors on August 20. The company was not identified, and the letter was not provided.
Those claims may become testable when the missing details appear. For now, the headlines are not enough.
Start With the Work
Nate B. Jones argues for starting with your best work and the thinking loop behind it. Then ask which model accelerates that loop.
Turbo Field Fair had the smallest reported process footprint here, but paid for it in prompt latency and SSD access. Four DGX Sparks combined compute fastest, but generated at one box’s speed. The M5 Ultra kept the model on one memory bus and reportedly generated faster, yet felt slower on long prompts.
The fastest result, the smallest memory footprint, and the best workflow fit were not always found in the same system.