DeepSeek-V4-Flash-Fast — local agent model for real terminal work

DeepSeek-V4-Flash-Fast is a custom agent build based on DeepSeek-V4-Flash: an open-weights Apache 2.0 MoE model with approximately 285B total parameters and 20B active parameters. It is configured for autonomous, tool-driven work on local hardware — especially coding, system administration and operational tasks performed through a real shell.

It does not introduce a new model architecture and is not claimed to have been trained from scratch. The architecture remains 100% DeepSeek-V4-Flash; its agent behaviour comes from prompt engineering, runtime configuration, sampling parameters and iteration against real agent harnesses.

What it’s for

• Agentic tasks in a native harness that executes real shell commands.
• Coding and system administration workflows that use tools rather than assumptions.
• System inspection: network scans, disk usage, Linux server inventory and similar operational tasks.
• Local deployment of a very large MoE model on high-end Apple Silicon or RTX-class hardware, without dedicated GPU clusters.

The model is designed to answer from real tool output. If the harness runs nmap, df, du or a remote server inventory command, the final answer should reflect the values returned by those commands rather than guessed hostnames, hardware details or numbers.

Model and deployment

The build goal is to make a 285B-class model useful on comparatively weaker hardware. With low-bit quantization, normal operation is possible using approximately 96 GB of unified memory, avoiding the need for expensive dedicated GPU packs.

Agent behaviour

The configuration was iterated using two models:

• Fable5 — selected for stable, timely tool calls and a direct “DO tasks” style.
• DeepSeek-V4 — used to iterate the agent loop and configuration against a real coding agent.

The base system prompt was tuned specifically for the DeepSeek Fast model. The result is an agent that acts first: it executes commands through the harness, checks real results and returns a grounded answer instead of producing a speculative explanation.

Prefill-focused performance

Agent harnesses impose a large prefill before the model starts working. For this reason, DeepSeek-V4-Flash-Fast prioritizes higher prefill throughput — prompt tokens per second — alongside usable inference performance.

This matters particularly for long system prompts, tool schemas, terminal transcripts and 128K-context workflows.

Runtime configuration

Designed for llama-server or llama-swap:

-t 16
-tb 16
-b 16384
-ub 16384
-c 131072
--temp 0.2
--top-p 0.95
--top-k 40
--repeat-penalty 1.15
--presence-penalty 1.0
--reasoning off
--jinja
--chat-template-file

The 131,072-token context setting provides a 128K working context. Low-temperature sampling and disabled reasoning are used to keep the agent focused on executing the task and reporting results grounded in tool output.

Tested agent harnesses

• OpenCode
• Pi coding agent
• Claude Code

Testing was performed on a GMKtec EVO x86 machine.

Measured behaviour

The model has completed the following end-to-end tasks using real tool output:

DeepSeek-V4-Flash-Fast is therefore intended not as a general conversational wrapper, but as a practical local agent: a large, low-bit MoE model configured to use a real environment, perform the requested work and answer from the results.

How it was made

This model was designed, built and tested with the help of Claude Fable5 and DeepSeek-V4 Pro. The idea is that the best coding models in the world should be able to create smaller models in their own image.

Its system prompt, parameters and context configuration come directly from that work: leading coding models preparing a local model that takes over directly on your desk.

License

Apache 2.0, inherited from the base DeepSeek-V4-Flash.

© 2026 — DeepSeek-V4-Flash-Fast