DeepSeek-V4-Flash-Fast
DeepSeek-V4-Flash-Fast is a custom agent build based on DeepSeek-V4-Flash (Open Weights, Apache-2.0, MoE, ~285B total / ~20B active). Weights in low-bit quantization for fully CPU-only deployment:
GGUF: DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf
Not trained from scratch — architecture is 100% DeepSeek-V4-Flash. Agent behaviour is prompt engineering + configuration.
Agent config origin
Tuned with only two models:
- Fable5 — stable, timely tool calls, "DO tasks" style.
- DeepSeek-V4 — iterating agent loop/config against a real coding agent.
The base system prompt was tuned for the DeepSeek Fast model.
Purpose — agentic tasks (harness)
A native backend that executes real shell commands and answers only from real tool output.
Build goal: a model for weaker machines — to work normally on a 285B model, only ~96 GB of unified memory is required. A big model runs on high-end Mac/RTX-class hardware without costly, dedicated GPU packs.
Prefill focus: agent harness imposes a large prefill. Work focused on increasing prefill throughput (prompt tokens/s).
Testing harnesses
- opencode
- pi coding agent
- claude code
Test machine: GMKtec EVO (x86).
Typical tasks (all passed — real results, no hallucination)
- Network scan
nmap+ interpretation. - Disk usage (
df/du) — real numbers. - Remote Linux server inventory — CPU/RAM/disk/OS.
- HTML5 Tetris — full working game.
- HTML5 solar system — animated orbits.
Runtime params (llama-server / llama-swap)
-t 16 -tb 16 -b 16384 -ub 16384 -c 131072 (128K)
--temp 0.2, --top-p 0.95, --top-k 40, --repeat-penalty 1.15, --presence-penalty 1.0, --reasoning off
--jinja --chat-template-file
(c) 2026 — DeepSeek-V4-Flash-Fast.