BREAKING
DeepSeek-V4-Flash Runs Locally
0
B
total params
0
B
active params
0
M
context tokens
Memory by quant variant
1-bit
92
2-bit
102
3-bit
135
4-bit
162
Q8
169
V4-Flash benchmark scores
MMLU-Pro
86.2
GPQA Diamond
88.1
Terminal Bench 2.0
56.9
Local run: gains and cautions
Upsides
praised
●
Lossless Q8 near 168GB
●
3-bit works on 110GB Mac
●
Tool-use and chat fixes
Cautions
noted
●
Q2/Q3 degrade long tasks
●
Cloud version still ahead
●
Large download times
Open weights, now on your desk
AI NEWS BLITZ
Unsloth has released quantized GGUFs that let DeepSeek-V4-Flash run fully offline.