bench3 · August 2026 · 16 tasks · 6 categories

Local LLMs step
into the arena.

Five cloud frontier models — DeepSeek V4 Flash & Pro, GPT-5.6 Sol, Grok 4.6, GLM 5.2 Vision (LunaRoute) — versus MTPLX, a free, open-source Mac app running Qwen 3.8 27B on a single laptop. Same harness, same tasks, same hidden graders. The local model holds its own.

75%
task pass rate
trial 1 · 12 / 16 tasks
$0.00
marginal cost
100% local & offline
4.24/5
judged output quality
blind rubric · creative + writing
1.7×
MTP decode speedup
vs plain autoregressive · up to 2.6×
Overview

Five models, one gauntlet

Every arm faced the identical bench3 harness: hidden-test grading for coding and agentic tasks, objective checkers for reasoning and tool use, and a blind rubric judge for creative and writing output. MTPLX ran the whole set on a laptop — the cloud arms on their vendors' GPUs.

Results

Where each model lands

Pass rate by category. Reasoning is where MTPLX shines — it solved the two hardest puzzles (r1_tiling, r4_catalan) that both DeepSeek arms missed. Tool use is its blind spot.

Per-task detail

Cloud arms: up to 5 trials. MTPLX: trial 1 (n=16). ⏱ = watchdog timeout, not a capability verdict.

Tokens / second

Speed, honestly measured

The harness records wall-clock time and token counts, so every run yields an effective tokens/sec. Cloud GPUs are fast; MTPLX is a laptop. But decode speed tells a different — and surprising — story.

MTP vs plain decoding

MTPLX accelerates generation with native MTP speculative decoding — the model's own multi-token-prediction heads draft ahead, and a batched verify pass checks them. Measured on this M5 Max (temp 0, 1,000 tokens):

MTP
22–30 tok/s
AR
13–18 tok/s

≈1.6–1.7× faster, up to 2.6× in a cool burst — identical output distribution.

Thermals beat samplers

The same 9,000-token prompt, same sampler, measured across a 2-hour session on the M5 Max:

A cool Mac decodes at 36–46 tok/s; after hours of sustained load it throttles to ~26. The app dashboard's 46.4 tok/s was a fresh, cool run.

Follow-up · 17 Aug 2026

The Qwen MLX challenge: a community 2.94×

After our benchmark, an open leaderboard (yukon.org/mlxfast) claimed Qwen 3.8 27B runs 193.4% faster on Apple Silicon via community-tuned native MTP kernels. We fetched the record submission (Layr-Labs/qwen-3.8-mtp-challenge), built it, and reproduced the measurement on this same M5 Max — 2.07× decode speedup with exact output parity (all tokens bit-identical to serial). The leaderboard's 2.94× is measured on a thermally gated, idle box with 8 hidden prompts; our machine was hot and GPU-contended during the run.

Verdict: confirmed. Native MTP decoding on Qwen 3.8 27B delivers a ~2× decode speedup on our hardware (11.2 → 23.3 tok/s, parity exact) — and the mechanism scales to ~2.9× under the leaderboard's gated conditions. Full methodology: MLXFAST_CONFIRMATION.md.
Cost

Total spend: $4.90 — and MTPLX was $0

Cloud arms burned real money across 250 runs. The local model produced its entire results column for free, on a laptop the user already owns.

Cost per passing task: MTPLX $0.0000 · DeepSeek Flash $0.0027 · DeepSeek Pro $0.0083 · Grok $0.0220 · GPT-5.6 Sol $0.1030. At scale, "free" compounds fast.
Setup

What MTPLX is, and what it ran on

MTPLX — native MTP on Apple Silicon

  • Native MTP speculative decoding. Qwen 3.8 ships MTP heads; MTPLX is one of the few runtimes that uses them — the model drafts several tokens ahead and verifies them in one batched forward pass.
  • Exact at any temperature. Leviathan–Chen probability-ratio acceptance with residual correction — output distribution provably unchanged, no second drafter model.
  • MLX + custom Metal kernels. Built for Apple Silicon; OpenAI & Anthropic-compatible local server (this benchmark used the /v1/chat/completions API).
  • Free & open source. Apache-2.0, built solo by @youssofal. Version 2.7.1, turbo profile, depth 3.

Hardware & model

LaptopMacBook Pro · Apple M5 Max
Memory64 GB unified
OSmacOS 26.5.2
RuntimeMTPLX 2.7.1 · MLX 0.32 · turbo MTP · depth 3
ModelQwen 3.8 27B Optimized Speed
Weights20.4 GB · 4-bit dynamic quant · MLX safetensors
Peak memory~22.4 GB of 64 GB
Context262,144 tokens
Method

How it was measured

Harness

  • 16 tasks · 6 categories: coding, agentic workflow, reasoning, tool use, creativity, writing — hardest-but-smallest probes, laddered to find where arms separate.
  • Objective, blind grading: hidden tests the model never sees; check.py validators; a separate judge model scores creative/writing on rubrics with no arm labels.
  • Identical tool surface: coding/agentic runs through the same agent harness; single-shot runs through one client with only wire params differing.

Caveats

  • MTPLX is trial 1 (n=16) vs 3–5 trials (n=50) for cloud arms — its 75% has wider error bars.
  • c2_perf = watchdog timeout (30 min), not a proven miss — it was mid-work, 29k tokens in, when killed.
  • Speed differs by temperature & thermals: a hot laptop is 40% slower than a cool one; the site shows both.
  • Costs use list prices at run time; MTPLX is $0 because it runs on your own hardware.