How long does a 64K prompt take on the 27B model?

About 110 seconds of prefill on Qwen3.8-27B at 5-bit. He wants that prefill, or time-to-first-token, reported separately from decode, and both numbers as the context grows.

What Mac Studio is he weighing, and what is the budget?

A desktop Mac Studio with 128 to 256 GB of unified memory: an M5 Max at 128 GB, or an M5 Ultra at 256 GB. Roughly $5,000 to $11,000 at current Apple prices, for sustained power and memory headroom.

What would make a tokens-per-second figure trustworthy?

The exact setup (hardware, framework version, quant, context length, sampling), time-to-first-token and decode reported separately, medians over several runs, and a script he can rerun.

What should a hardware-fit site tell him?

Whether the model fits at the context he needs, agent-style latency for prefill and multi-step tool calls, which frameworks run the model correctly, and how much each quant hurts real tasks. Full-context memory math should include the KV cache.

May we quote Xijing (Thomas) Wang by name?

Yes. He wrote that we are welcome to quote him by name. The site he pointed at is https://xthomaswang.github.io and the handle is @xthomaswang.

Can I test open models in the browser before buying a Mac Studio?

Yes. You can test open-weight models side by side in your browser at https://localllmrig.com/. Qwen3.8-27B, his daily model, is at https://localllmrig.com/can-i-run/qwen3.8-27b.

Xijing (Thomas) Wang's MacBook Pro M5 Max 128 GB

A MacBook Pro M5 Max with 128 GB of unified memory. About 107 GB of that is usable by the GPU. Qwen3.8-27B at 5-bit, served by oMLX with MTP speculative decoding, is the daily backend for Claude Code, Hermes, and his own agent.

Xijing (Thomas) Wang (@xthomaswang), quoted by name with written permission. 27 Oct 2026. Self-reported. Not a localLLMrig measurement.

Test Gemma 4 31B, Recursive Flint and Nemotron 3.5 Super VL side by side in your browser, free.

xthomaswang.github.io. github.com/xthomaswang.

The laptop

Everything runs on the MacBook Pro M5 Max, 128 GB. The daily model is Qwen3.8-27B at 5-bit, served by oMLX with MTP speculative decoding. That server is the backend for Claude Code, Hermes, and his own agent. He is also testing Nanbeige4.3-33B and Qwen3.8-Flash-Next, a mixture-of-experts model of about 100 GB. Smaller embedding models on the same machine are Qwen3-Embedding-0.6B and bge-m3.

Two limits show up in real use. The GPU can use about 107 GB of the 128 GB. On a 96 W power adapter the battery drains during a long GPU run.

Long context

A 64,000-token prompt takes about 110 seconds of prefill on the 27B model. He wants time-to-first-token, which is the prefill, reported separately from decode, and he wants both as the context grows. He also wants the memory math at full context, including the KV cache, and agent-style latency: prefill plus multi-step tool calls, not decode tokens per second on their own.

New model hardware requirements and price updates by email. We email a link to confirm the address. Nothing is sent until you click it. Privacy notice.

The next machine

A Mac Studio with 128 to 256 GB of unified memory. That is an M5 Max at 128 GB, or an M5 Ultra at 256 GB. The budget is roughly $5,000 to $11,000 at current Apple prices. The desktop is for sustained power and for memory headroom the laptop does not have.

Where he looks before buying

Official docs and model cards: Apple MLX notes, Hugging Face model cards, and framework docs. Also posts on X from people who measured the machine themselves. What is usually missing is prefill time at long context, decode speed as the context grows, and memory at full context including the KV cache.

What would make a speed figure trustworthy

The exact setup: hardware, framework version, quant, context length, and sampling. Time-to-first-token and decode reported separately. Medians over several runs. A script he can rerun.

What he wishes a hardware site told him

Whether the model fits at the context he needs. Agent-style latency, meaning prefill and multi-step tool calls, not only decode tokens per second. Which frameworks run the model correctly. How much each quant hurts real tasks.

On our list: Can I run Qwen3.8-27B?, Can I run Nanbeige4.3-33B?, Can I run Qwen3.8-Flash-Next?, Mac Studio M5 Max 128GB, Mac Studio M5 Ultra 256GB.

The numbers on this page were measured by the owner on that hardware. They are not a localLLMrig benchmark, and they do not change the fit scores in the calculator.

Questions

What does Xijing (Thomas) Wang run locally, and on what hardware?: A MacBook Pro M5 Max with 128 GB of unified memory, about 107 GB of it usable by the GPU. Qwen3.8-27B at 5-bit, served by oMLX with MTP speculative decoding, is the daily backend for Claude Code, Hermes, and his own agent. He is testing Nanbeige4.3-33B and Qwen3.8-Flash-Next (about 100 GB). Embedding models: Qwen3-Embedding-0.6B and bge-m3. Self-reported, 26 October 2026.
How long does a 64K prompt take on the 27B model?: About 110 seconds of prefill on Qwen3.8-27B at 5-bit. He wants that prefill, or time-to-first-token, reported separately from decode, and both numbers as the context grows.
What limits does the MacBook Pro hit?: GPU-usable memory is about 107 GB of the 128 GB. On a 96 W power adapter the battery drains under a long GPU run.
What Mac Studio is he weighing, and what is the budget?: A desktop Mac Studio with 128 to 256 GB of unified memory: an M5 Max at 128 GB, or an M5 Ultra at 256 GB. Roughly $5,000 to $11,000 at current Apple prices, for sustained power and memory headroom.
What would make a tokens-per-second figure trustworthy?: The exact setup (hardware, framework version, quant, context length, sampling), time-to-first-token and decode reported separately, medians over several runs, and a script he can rerun.
What should a hardware-fit site tell him?: Whether the model fits at the context he needs, agent-style latency for prefill and multi-step tool calls, which frameworks run the model correctly, and how much each quant hurts real tasks. Full-context memory math should include the KV cache.
May we quote Xijing (Thomas) Wang by name?: Yes. He wrote that we are welcome to quote him by name. The site he pointed at is https://xthomaswang.github.io and the handle is @xthomaswang.
Can I test open models in the browser before buying a Mac Studio?: Yes. You can test open-weight models side by side in your browser at https://localllmrig.com/. Qwen3.8-27B, his daily model, is at https://localllmrig.com/can-i-run/qwen3.8-27b.

Find hardware · All builder interviews · Send us your llama-bench result