Skip to content

Latest commit

 

History

History
90 lines (59 loc) · 3.9 KB

File metadata and controls

90 lines (59 loc) · 3.9 KB

Communication Summary

The Honest Story

We don't make things faster. We make impossible things possible, and tight things comfortable.

Three Real Use Cases

1. Model barely fits your RAM (80-100%) — the sweet spot

Your 36GB Mac runs a 30GB model. macOS memory pressure kicks in — compression, swapping, GPU stalls. You lose 25-50% of your speed without knowing why.

We shrink the model's memory footprint by 25% using mixed precision (hot experts at 4-bit, cold at 2-bit). The model now effectively uses 22GB instead of 30GB. Memory pressure disappears. Full speed restored.

35GB model on 36GB Mac (97% RAM used):
  Without us:  8.3 tok/s  ████████          (OS thrashing, -40%)
  With us:    13.9 tok/s  ██████████████    (+67% — pressure eliminated)

2. Model too big for RAM (100-500%) — makes it possible

A 209GB model on a 36GB Mac. Without us: OOM crash, doesn't run at all.

With us: runs at 3-7 tok/s by streaming from SSD with intelligent caching. That's ~1 word per 0.3 seconds — slow, but it works. Good enough for:

  • Code review (ask a question, go make coffee, get a GPT-4-quality answer)
  • Document analysis (submit a doc, wait 30 seconds for a thorough analysis)
  • Research (quality matters more than speed)
209GB model on 36GB Mac:
  Without us:  CRASH (doesn't run)
  With us:     3.7 tok/s  ███████████  (slow but working)

3. Model fits easily (<70% RAM) — no help needed

We add zero value here. Pure MLX runs at full speed. We don't pretend otherwise.

What's Actually New

Nobody else combines all of these on Apple Silicon:

  • Task-aware caching (different tasks use different 30% of the model)
  • Adaptive memory management (never harms your other apps)
  • Mixed precision per-expert (hot=4bit, cold=2bit, decided at runtime)
  • C GCD acceleration (Apple's native dispatch, 8x faster than Python)
  • Live topic change detection (re-caches in <2 seconds)

Similar projects exist (mlx-moe, PowerInfer, flash-moe) but each solves only part of the problem. We integrate the full stack.

Key Numbers (All Measured on M3 Max 36GB)

What Number How Measured
MLX baseline 115.9 tok/s Qwen MoE, 3 runs averaged
Cache hit rate 85.4% 24 layers, 60 experts, LCP eviction
Cache overhead 3.9ms/token Fits inside 9.1ms GPU time = FREE
Mixed precision 1.80x smaller MSE 0.000059, negligible quality loss
Topic switch 92% cache swap Adaptive profiler detects in <2s
Memory-safe Auto-adjusts Monitors macOS pressure, never harms user
Tests 59 passing Covers all modules

Social Posts

Twitter (honest version)

Running a 200GB AI model on a 36GB MacBook sounds impossible. It is — without smart caching.

We built MLX-Flash. It streams model weights from SSD with 85% cache hit rate. Result: model runs at 3-4 tok/s instead of crashing.

That's slow. But "slow" beats "doesn't run at all."

The surprising win: models that BARELY fit (90-100% of RAM). macOS memory pressure causes 25-50% speed loss. Our mixed precision shrinks the footprint 25%, eliminating pressure entirely. Full speed restored.

github.com/szibis/MLX-Flash

LinkedIn (professional)

Running models that "almost fit" your hardware — the overlooked problem

Everyone talks about running huge models on small devices. But there's a quieter problem: models that technically fit in RAM but perform terribly because macOS memory pressure causes GPU stalls.

We built MLX-Flash to solve three problems:

  1. Barely fits: 35GB model on 36GB Mac → +67% speed (pressure eliminated)
  2. Doesn't fit: 209GB model on 36GB Mac → 3.7 tok/s (impossible → possible)
  3. Comfort zone: adaptive memory management that never harms your other apps

Built on 60+ research papers, measured on real hardware, 59 tests passing.

Open source: github.com/szibis/MLX-Flash