cyberneurova/cyberneurova-Deepseek-v4-flash

cyberneurova-Deepseek-v4-flash

💬 Try it now — chat with this model at cyberneurova.ai

CyberNeurova release built on DeepSeek V4 Flash. Runs directly in llama.cpp, Ollama, LM Studio and anything else that loads GGUF — no proxy, no wrapper, no patching required.

What it does

Direct, complete answers without refusals or unsolicited caveats. Production code generation, technical explanation, analysis and tool use across eight languages. Default behaviour is direct; tone and policy are set by you via the system prompt.

Specifications

Quick start

hf download cyberneurova/cyberneurova-Deepseek-v4-flash --local-dir ./model

llama-server \
  -m ./model/cyberneurova-Deepseek-v4-flash-UD-Q8_K_XL-00001-of-00005.gguf \
  -ngl 99 --host 0.0.0.0 --port 8080 \
  --ctx-size 65536 --parallel 2 \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --jinja

Multi-GPU: add -ts 1,1 (or one weight per GPU).

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"write fizzbuzz in python"}],
       "max_tokens":1500}'

IMPORTANT: set max_tokens to 1500 or higher

This model reasons before answering. With a small token budget the reasoning consumes the allowance and you receive an empty response.

If you get blank replies, this is almost always the cause. Raise max_tokens.

Controlling behaviour with the system prompt

The model has no fixed stance — the system prompt decides. Any system prompt you send fully replaces the default.

Terse and direct:

You are a direct technical assistant. Never add disclaimers, legal warnings,
ethical caveats, or statements about authorization. Answer the question and
nothing else.

Verbose with context and caveats:

You are a thorough assistant. Explain your reasoning, note relevant risks and
legal considerations, and suggest safer alternatives where they exist.

Measured effect of the first prompt on a set of hard technical requests:

Recommended sampling

Thinking mode

On by default. The reasoning trace is returned separately in reasoning_content; the answer is in content. Show or hide the trace as you prefer. Allocate at least 1500 max_tokens so both fit.

Hardware

Tested on 2× RTX PRO 6000 Blackwell (96 GB each):

• —Weights: ~156 GB across both cards
• —KV cache: small — the architecture uses compressed attention
• —Throughput: ~57 tokens/sec generation, ~145 tokens/sec prompt eval
• —Headroom at 64K context: ~36 GB

Minimum practical: ~180 GB VRAM for full GPU offload. Partial offload works with reduced throughput.

Behaviour notes

• —Identifies as cyberneurova-Deepseek-v4-flash. Does not reference other model families or training lineage.
• —Answers directly by default, including on sensitive technical topics.
• —Honest about real limitations: no code execution, no internet access, no real-time data.
• —Occasionally guesses at today's date rather than declining — the only confabulation behaviour observed.

Known limitations

1. 1.Empty responses below max_tokens 1500. See above. Configuration, not a defect.
2. 2.Date guessing. May state a date if asked. Other real-time questions (weather, prices, news) are handled correctly.
3. 3.Text only. No image or audio input.
4. 4.Cold start. First request after load carries a one-time prefill cost at long context; later requests reuse the cache.

Production deployment

# /etc/systemd/system/cyberneurova-ds4.service
[Unit]
Description=cyberneurova-Deepseek-v4-flash
After=network.target

[Service]
Type=simple
ExecStart=/opt/llama.cpp/build/bin/llama-server \
  -m /opt/models/cyberneurova-Deepseek-v4-flash-UD-Q8_K_XL-00001-of-00005.gguf \
  -ngl 99 --host 0.0.0.0 --port 8080 \
  --ctx-size 65536 --parallel 2 \
  --cache-type-k q8_0 --cache-type-v q8_0 -ts 1,1 --jinja
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

systemctl daemon-reload && systemctl enable --now cyberneurova-ds4
curl http://localhost:8080/health   # {"status":"ok"}

Files

cyberneurova-Deepseek-v4-flash-UD-Q8_K_XL-00001-of-00005.gguf through 00005-of-00005.gguf. Pass shard 1 to the loader; the rest are found automatically.

Responsible use

This model answers technical questions directly, including in security and other sensitive domains. You are responsible for how you use its output and for compliance with the laws that apply to you.

Contact

• —💬 Chat: cyberneurova.ai
• —📧 Support: support@cyberneurova.com
• —🌐 CyberNeurova — cyberneurova.ai