DeepSeek V4.1 Flash 1017
GSQ-RCO IQ1_M backbone + shared Engram IQ4_NL + mmproj · ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
queued
| Publisher | ISTA-DASLab |
|---|---|
| Modality | text+vision LLM |
| Parameters | 763.2B total / unknown active 763.2B per HF safetensors count, including 196.6B of Engram lookup tables (ISTA-DASLab card). 384 routed experts, 6 per token (config.json). Active parameter count is not stated on the official card: unknown. Source: model card. |
| Format | GSQ-RCO IQ1_M backbone + shared Engram IQ4_NL + mmproj |
| Size on disk | 216.34 GiB (232,288,767,636 bytes) |
| Estimated memory to run | ~130.8 GB (estimate; see method) |
| Native context | 1,048,576 tokens (config/card) |
| Engine | llama.cpp (build from Oct 27 2026 or later) · recipe |
| License | mit MIT: commercial OK. |
| Archive folder | official |
| Notes | Backbone 120.72 GB must be resident; the 110.60 GB Engram file can stay memory-mapped on SSD (ISTA card). Engram must be linked as shard 6. |
Where it runs
| Machine | Verdict | Budget used |
|---|---|---|
| Bosgame M5 (x1) | No | ~108-110 GB GPU-usable on Linux with GTT raised (96 GB cap on Windows). Fit budget used here: 108 GB. |
| Bosgame M5 (x2, RPC) | Fits | ~200-225 GB usable across two nodes. Fit budget used here: 210 GB; 210-225 GB = tight. |
| MacBook Pro (M4 Max) | No | macOS caps GPU wired memory by default (~75% = ~27 GB). Raising iogpu.wired_limit_mb to ~32 GB is possible but leaves little for the OS. |
| Mac mini (M4) | No | ~2/3 of 16 GB usable by the GPU by default. Already busy as build box + Hermes/Ollama. |
| omarchy (AMD desktop) | No | 7 GB of the 8 GB VRAM for weights. Offload capacity depends on system RAM, which is unknown today; after the 128 GB upgrade ~110 GB of MoE experts could sit in RAM. |
| Ugreen DXP2800 NAS | No | Storage host. Already runs ollama/open-webui containers; only tiny models (e.g. Whisper on CPU) are realistic. |
| MacBook Neo | Unknown | Specs unknown: machine was offline/unreachable on Nov 14 2026. |
Token/s = theoretical ceiling from memory bandwidth and active weight bytes; not a benchmark. Nothing on this page is a benchmark score.