Skip to content
llm-speed
← raw run
wrapped·Apr 28, 2026
/r/r_l_v1-zq_qaz
your run

DeepSeek-Coder-V2-Lite-Instruct-4bit

on M3 Ultra
168
tok/sdecode
M3 Ultra (60-core GPU) + 96GB unified
rank in tier
2/20M3 Ultra runstop 10%
best workload
chat-short
where the rig flew
slowest workload
—
single-workload run
backend
mlxMLX is the fastest backend on Apple Silicon for dense models.
faster than
  • gpt-4o-mini on M3 Pro158 tok/s
  • llama-4-maverick on M3 Pro126 tok/s
  • gpt-4o on M3 Pro83.8 tok/s
share
168 tok/s on M3 Ultra running DeepSeek-Coder-V2-Lite-Instruct-4bit — top 10% of M3 Ultra runs. Signed benchmark via llm-speed.
https://llm-speed.com/wrapped/r_l_v1-zq_qaz
X / TwitterBlueskyMastodonRedditraw run

This is a shareable view of one signed llm-speed suite-suite-v1 run. Numbers are not edited. The raw record has every workload, the public key, and the fingerprint hash.