M6 Launch Day: 9.1 vs 10.6 tok/s — Should You Upgrade 16GB to 32GB for Local LLMs?

Four Benchmarks Converge on the Same Bandwidth Ceiling

On launch day (Nov 6), the 14-inch M6 MacBook Pro starts at ¥15,999. The key metric isn't a synthetic benchmark but 9.1 tok/s : a 32GB M6 running Qwen3.8-27B 4-bit generates only ~9 tokens per second. Using the memory bandwidth formula — 170 GB/s ÷ ~16 GB model weight — the theoretical maximum is 10.6 tok/s . The measured 9.1 tok/s already consumes 86% of that ceiling, indicating a hard physical limit, not a software optimization gap.

Four independent tests from different publishers, machines, and software all cluster between 9–10 tok/s:

Community test: 32GB M6 Mac mini, Qwen3.8-27B 4-bit, 1K context → 9.1 tok/s ; at 32K context drops to 7.9 tok/s (13% decline), peak memory 18.8 GB, 100 tokens take ~11 seconds.

ComputerBase (via LM Studio 0.4.24, Q4_K_M): 32GB M6 Mac mini → 9.5 tok/s ; same machine runs 14B at 17.8 tok/s.

Launch review: 32GB M6 MacBook Pro, 27B model, ~23 GB memory → 9.7 tok/s .

少数派 test: 32GB M6 MBP in LM Studio → ~10 tok/s ; the same article shows a 16GB M5 blocked from loading 27B, forced to Gemma 4 12B (M6 19 tok/s vs M5 16 tok/s).

Because variables are messy yet results are tightly bounded, the conclusion is that software differences are flattened by the physical bandwidth limit ; switching inference frameworks cannot recover the missing 1.5 tok/s.

Cross-check with smaller models on 16GB machines (153 GB/s bandwidth): Qwen3.5 9B runs at 34.4 tok/s, 14B at 15.3 tok/s. Reverse-calculating effective weight read per token yields ~4.4 GB and ~10 GB respectively, matching the approximate 4-bit file sizes. The formula target tok/s ≈ bandwidth ÷ weight size holds as an upper bound; actual throughput is further reduced by quantization format, cache hits, and engine implementation. At 27B dense, headroom disappears.

"Only 8 Models" Is a Software Filter, Not a Hardware Wall

Geek Culture's launch review showed LM Studio's Bionic recommending only 8 models (max 7B) for 16GB M6, versus 20+ models (up to 170 GB) on a 128GB Mac Studio. The review explicitly states this does not reflect the chip's capability — it's a client-side memory-based recommendation list.

Public benchmarks contradict the list: 16GB M6 runs 9B at 34.4 tok/s, 14B at 15.3 tok/s, and even a 16B MoE at 79.4 tok/s. Manually downloading GGUF files from Hugging Face allows many more models. The real constraints are capacity and experience : 27B 4-bit peaks at 15.8 GB (1K context) and 18.8 GB (32K context), exceeding 16GB; running browser, IDE, and model service simultaneously triggers swapping.

What the Dual Neural Engine Actually Accelerates

Apple's press release carefully distinguishes workloads: the "13× faster" claim for 14-inch MBP compares prompt processing (prefill / first token) against 13-inch M1; versus 14-inch M5 it's only 1.3× . The "25% faster" device-side AI example is LLM prompt processing. The "1.9×" dual neural engine example covers image generation, super-resolution, and quality enhancement. Mac mini's "4.8×" also refers to LM Studio prompt handling.

Measured split: on 9B 4-bit, M4 prefill 210 tok/s → M6 742 tok/s (3.5×), but decode M4 18 tok/s → M6 26.9 tok/s (1.5×). Independent dual-engine testing shows compute-bound loads scale ~1.92× per engine, while weight-read-bound loads only ~1.4× because both engines share the same memory bandwidth.

Toolchain reality: MLX, llama.cpp, Ollama use GPU/CPU; the NPU requires Core ML. Downloaded MLX models for chat rarely engage the dual neural engine. However, for vision tasks the ANE delivers: ResNet-50 batch 16 is 2.33× faster, OCR text detection 2.69× faster, Flux.1 Schnell generates in 35 seconds vs 93 seconds on M4. Buy for on-device vision, super-resolution, Apple Intelligence — not for doubling LLM token generation.

Three Workarounds for the Bandwidth Wall

The statement "M6's AI narrative is welded shut by bandwidth" is true only for dense model native autoregressive decode , not all local AI. Three bypasses:

Switch to MoE : On 32GB, Qwen3.6-35B-A3B 4-bit reaches 63–69 tok/s; on 16GB, 16B MoE hits 79.4 tok/s. Each token reads only active experts, shrinking the denominator in the bandwidth formula.

Speculative Decoding / MTP : One test lifted 27B from 8.6 to ~52 tok/s (~6×, estimated). On base M5, Qwen3.5 9B: llama.cpp 22 tok/s, MLX 25.1 tok/s, Uzu with speculative decoding 92.1 tok/s (4.2×). Gains depend heavily on draft acceptance rate; treat as experimental, not guaranteed.

Change Inference Engine : Same hardware, different software yields gaps larger than a generational chip upgrade.

These are single-machine, non-standardized tests; speculative decoding gains are highly variable. They prove 9.1 tok/s is the ceiling for "vanilla 27B dense load", not the machine's absolute speed limit .

Buying Guide: 16GB, 24GB, or 32GB?

Three user profiles:

Just want a local assistant : 16GB suffices, but don't buy the MBP. The M6 Mac mini starts at ¥6,999 — the ~¥9,000 premium buys screen, battery, portability, not local AI performance.

Serious local model work : Minimum 24GB (¥17,499). The extra ¥1,500 unlocks not only 8GB more capacity but also upgrades bandwidth from 153 GB/s to 170 GB/s — the best value tier.

Targeting 27B+ models : 32GB fits 27B dense, but 9–10 tok/s is "quality tier, not daily driver". At this price, adding more reaches M5 Pro's 307 GB/s. For 70B or long-chain agents, start at M5 Pro 64GB (~¥24,249 estimated).

Pre-order checklist: (1) List target models, apply weight size × target tok/s ≤ bandwidth to estimate feasibility. (2) Check 32K context peak memory and system headroom, not just "can it load". (3) Test the same model with MLX, llama.cpp, and speculative decoding before deciding to spend more.

Disclaimer : Benchmark numbers come from various media/community sources with differing hardware, software versions, and context lengths — not directly comparable. Apple figures are vendor "up to" claims. All reverse-calculated values and the ¥24,249 estimate are approximations.

References: Launch-day benchmark aggregation (什么值得买) · Dual neural engine teardown (什么值得买) · 少数派 M6 hands-on · ComputerBase data · Apple China online store.