This week's 3.5-hour real-world test on Reddit left us with a contradictory takeaway: a 128GB MacBook Pro M5 Max can sustain a 350,000-character context window (roughly three novellas' worth), but past the 100,000-character mark the model starts treating the user's words as its own output and continues answering by mixing them in.
What this is
The test used Qwen3.8-Flash-Next — a preview-grade small model from Alibaba's Tongyi Qianwen. After 2-bit quantization (crushing parameter precision to the floor to save memory), it weighs around 79GB, slotting neatly into the 128GB unified memory. The context window was stretched from 260K to 350K characters via YaRN (a technique that forcibly extends a model's "native" context length). Key numbers we want you to hold onto: peak first-pass input ingestion hit 1,561 characters/second; generation ran at 30+ characters/second. By the time the conversation reached 170K characters, generation had dropped to 11.5 characters/second.
Industry view
The optimists will stress: 128GB unified memory, Metal acceleration, and YaRN together prove consumer-grade hardware can now deliver near-enterprise performance locally. For data-sensitive settings like law firms and hospitals, we read this as a real signal — there's no longer a hard requirement to ship client documents to public clouds.
But we think the objections deserve a hearing too. The tester explicitly flagged "role confusion" past 100K characters — the model treats the user's words as its own output and keeps fabricating from there. This isn't isolated: 2-bit quantization inherently bleeds model precision (lower precision compresses parameters harder, and the model's "comprehension" takes a real hit), and layering an immature preview-grade long-context capability on top magnifies the problem. In short: the hardware bar has collapsed, but a deep quality chasm still separates "runnable" from "actually useful." One more thing worth flagging: Apple's memory-upcharge strategy makes 128GB and 192GB configurations notably more expensive, so ordinary users looking to run local LLMs aren't getting a cheap ride.
Impact on regular people
For enterprise IT: High-memory Mac workstations are now worth evaluating for local AI deployment, especially where data compliance is non-negotiable.
For working professionals: A loaded Mac can now power offline "ask the whole book" tools, but the 100K-character quality cliff is a real risk — don't migrate core workflows just yet.
For the consumer market: By pushing unified memory to 128GB and 192GB, Apple has quietly raised the "AI PC" hardware bar — a product-strategy signal worth watching.