How will people run DeepSeek Coder v2 236B locally by 2025?
Plus
0
Jan 2
1D
1W
1M
ALL
50%
Lobotomy levels of quantization (e.g. Q2_K)
50%
Unified memory (e.g. M3 Ultra, Mac Studios)
50%
Non-GPU main memory (e.g. AMD EPYC with 512GB DDR5)
50%
Gaming GPUs in one motherboard (e.g. 4090s)
50%
Tall clustering (e.g. Mac Studios over Thunderbolt)
50%
Wide clustering (e.g. Petals)
50%
HBM FPGA/ASIC dark horse (e.g. AMD rains Versal chips like manna)
DeepSeek Coder v2 is arguably a frontier model from a stellar Chinese team. It has cutting edge efficiency tricks, improved reasoning thanks to extensive code pretraining, and what looks like an excellent math corpus.
It's also an absolute chonker despite being MoE.
What's going to be the path for GPU poors needing to run LLMs in "ITAR mode" without building their own private datacenter or buying NVIDIA racks?
Resolves to as many options that apply, based on public information. If there is no public information resolves based on vibes and a DIY pricing spreadsheet.
(Chinese overlords variant of Llama 3 405B question in case Meta gets cold feet.)
Get Ṁ1,000 play money
Related questions
Related questions
Will DeepSeek's next reasoning model be called R3?
5% chance
When will DeepSeek release R2?
Will there be an open replication of DeepSeek v3 for <$10m?
55% chance
When will DeepSeek release V4?
Will DeepSeek be banned in US before 2027?
18% chance
Did DeepSeek use a cluster of more than 5,000 H100s in 2024?
26% chance
Will DeepSeek's next reasoning model be open-sourced?
86% chance
will DeepSeek become a closed AI lab by EOY?
25% chance
How much did DeepSeek-V3 cost to train?
Will ANY DeepSeek model cause or materially enable a catastrophic risk by 2027?
12% chance