Strata 0.2.4 runs DeepSeek V4.1 Flash locally: 94 tok/s on a 12GB RTX 5070
Strata's terminal build bundles V4.1 Flash and serves a local OpenAI/Anthropic API; on a 12GB RTX 5070 with 64GB RAM, Q2_0 hits 94 tok/s and IQ3_S 53 tok/s, with prompt intake around 2,650 tok/s.
DeepSeek V4.1-Flash now runs inside 64GB of system RAM on a gaming PC. Strata (MIT) ships it in its Terminal build from 0.2.4 and serves an OpenAI/Anthropic API on localhost, so Claude Code and Codex can point at it. What the repo commits to: 32GB of RAM plus VRAM is enough to start, 64GB runs every size. On a 12GB RTX 5070 with 64GB of RAM, Q2_0 answers at 94 tokens/s, IQ3_S at 53, and prompt intake runs ~2,650 tokens/s. Q2 and IQ3 are 2-3 bit weights, so the open question is whether it holds a repo together. what would you test it on first?
This post quotes another post: “DeepSeek V4.1 Flash runs on 64GB RAM via Strata and GSQ-RCO quants”