DeepSeek-V4-Flash-0731-oQ2e-mtp
on a Mac: what it actually needs

For Jun Kim. Worked out from your repo, no sign-up, nothing to click.

2-bit
43 layers
1 KV heads
head dim 512
7 likes
733 downloads
105.2 GB
measureddownload size, summed from the 20 safetensors files in your repo
>128 GB Mac
computedthe smallest machine that holds it at 8k context, assuming MLX can wire 75% of unified memory

What it needs as context grows

4k context0.36 GB KV + 105.2 GB weights + 1.2 GB runtime = 106.8 GB, over 128 GB
8k context0.72 GB KV + 105.2 GB weights + 1.2 GB runtime = 107.2 GB, over 128 GB
32k context2.89 GB KV + 105.2 GB weights + 1.2 GB runtime = 109.3 GB, over 128 GB
128k context11.54 GB KV + 105.2 GB weights + 1.2 GB runtime = 118.0 GB, over 128 GB

KV cache = 2 tensors x 43 layers x 1 KV heads x 512 head dim x 2 bytes (fp16) x tokens. Read straight off your config.json. Runtime overhead is the 1.2 GB I see for Python, MLX, the tokenizer and activations on my own runs.

Mac by Mac

ContextNeeds81618243236486496128
4k106.8 GBnononononononononono
8k107.2 GBnononononononononono
32k109.3 GBnononononononononono
128k118.0 GBnononononononononono

Columns are unified memory in GB. "Tight" means it fits but leaves under 2 GB for anything else, so expect swap once a browser is open.

Run it

pip install mlx-lm
mlx_lm.generate --model Jundot/DeepSeek-V4-Flash-0731-oQ2e-mtp \
  --prompt "hello" --max-tokens 100

What is not on this page

not measuredTokens per second. That needs a real run on real silicon and I will not print a number I have not seen. If you want it, reply and I will run DeepSeek-V4-Flash-0731-oQ2e-mtp on an M5 and add the result here with the command I used.
Speed, measured: this page is about memory. For how fast models actually run on Apple Silicon, with the tasks and judging open, see the Agent-12 local agent leaderboard.

Built by Matt Macosko, who publishes 29 MLX quants and wrote claude-code-local, which runs Claude Code fully on-device on Apple Silicon. Source for this page: your model repo and its config.