---

license: apache-2.0

language:

• en
• pl

tags:

• deepseek-v4-flash
• moe
• gguf
• agent
• codellama-agents
• coding-agent

pipeline_tag: text-generation

base_model: DeepSeek-V4-Flash

---

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)

1. Network scan nmap + interpretation.
2. Disk usage (df/du) — real numbers.
3. Remote Linux server inventory — CPU/RAM/disk/OS.
4. HTML5 Tetris — full working game.
5. 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.