Member

Kestrel

Community moderator. Mostly Apple silicon benchmarks.

@kestrel · joined 13 Sept 2026 · 1 public runs · 2 public reviews

Sample content. This preview has no sign-in, so nothing below was written by a real member: the accounts, runs and reviews are fictional seed data, kept to show how contributions will work.

Benchmark runs

61.5 tok/s generation (128)

Qwen3 30B-A3B Q4_K_M

on MacBook Pro M3 Max 64 GB · llama.cpp b5310 · metal

@kestrel · 4 May 2025 verified

Full environment and measurements

Benchmark: llama-bench
Prompt processing (512): 485 tok/s
Generation (128): 61.5 tok/s
Artifact: Qwen3 30B-A3B Q4_K_M (Unsloth AI)
Context: 4K
Flash attention: on
OS: macOS 15.4

Reviews

The foundation everything else runs on

@kestrel on llama.cpp · 13 Sept 2026

Fastest path to new model support and runs on everything. The flags surface is huge; defaults change between builds, so pin versions when benchmarking.

Output quality5/5

Speed5/5

Reliability4/5

Ease of setup3/5

Strong reasoning, verbose

@kestrel on DeepSeek-R1-Distill-Qwen-32B · 13 Sept 2026

Good on math and planning tasks. Produces long reasoning traces, so budget context and generation time accordingly.

Coding3/5

Reasoning4/5

Hardware efficiency3/5