Akan4ik

DeepSeek-V4-Flash-0731-FFT-BF16-Cyber

Introduction

This model is a Hermes-native agentic pentest profile based on DeepSeek-V4-Flash-FFT-model for supervised, authorised penetration tests — from reconnaissance through safe exploit validation to reporting. It uses Hermes' built-in toolsets, three routed skills, and the required violin-guard plugin at the target-execution boundary. The standalone CLI supports release checks, diagnostics, and administrative recovery; target commands run through the plugin. Violin adds no profile-specific credentials and inherits the provider and tool backends already configured in Hermes. Notes:

For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95.
† DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.

Features

Chat

python
from encoding_dsv4 import encode_messages, parse_message_from_completion_text

messages = [
    {"role": "user", "content": "hello"},
    {"role": "assistant", "content": "Hello! I am DeepSeek.", "reasoning_content": "thinking..."},
    {"role": "user", "content": "1+1=?"}
]

# messages -> string
prompt = encode_messages(messages, thinking_mode="thinking", reasoning_effort="max")

# string -> tokens
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-0731")
tokens = tokenizer.encode(prompt)

Local tests

How to Run Locally

For local deployment, I recommend setting the sampling parameters to temperature = 1.0, with top_p = 0.95 for agentic scenarios and top_p = 1.0 otherwise. For the high and max reasoning effort levels, we recommend a maximum output length of 384K tokens.

License

This repository and the model weights are licensed under the MIT License.

Hermes Agent >= 0.18.0 — installed and on your PATH
Hermes provider configured — Violin inherits your normal Hermes provider/model
Kali Linux or Parrot OS — the primary execution environments; Docker Kali is the supported fallback when the host lacks pentest tools
Optional web/browser backend — required only for Hermes web or browser capabilities; Violin does not add separate API credentials

bash
hermes profile use violin

Engagement Workflow

Table with columns: Phase, Action, Safety Gate
Phase	Action	Safety Gate
1. Scoping	9 questions via `clarify`	User approval
2. Reconnaissance	Passive OSINT → tech detection → active scanning	Guard + approval
3. Vuln Research	CVE lookup, exploit search, attack surface analysis	Guard check
4. Exploitation	Safe PoC validation per vulnerability class	Guard + user approval
5. Reporting	Evidence compilation, CVSS scoring, remediation

Enabled Toolsets

11 toolsets configured in config.yaml (platform_toolsets.cli): 10 built-in — terminal, web, browser, file, code_execution, skills, todo, clarify, delegation, vision — plus the violin_guard guard-plugin toolset.

Conversation & Memory Isolation

memory.memory_enabled: false — no global memory recall/write
memory.user_profile_enabled: false — no global user profile access
Engagement continuity lives in project files (scope docs, evidence, reports)
Keep one Hermes conversation per engagement; after compression, resume in that conversation from $ENG_DIR/state/

Authorised testing only — no probing before scoping is complete
Approval gates — scope, active recon, and exploitation each require explicit user approval
Guard check — every target-touching command validated through violin_exec or another typed guard tool. violin_exec has no binary allowlist, so any installed non-interactive Kali/Parrot CLI tool can target the explicit in-scope host while the same scope, phase, PTT, hypothesis, history, evidence, timeout, and sync gates remain active. Violin's pre_tool_call plugin hook generically blocks target literals in raw terminal commands instead of maintaining a partial tool-name list. The CLI exposes the same check for diagnostics (exit 0=allowed, 1=blocked, 2=review)
Non-destructive by default — exploitation limited to safe, reproducible PoC
Evidence-first — every finding backed by reproducible tool output, screenshots, request/response pairs
Exploit-first validation — no hypothesis advances to Validated without a verification command
Stateful recovery — phase summaries and checkpoints restore the current engagement after context compression without starting a new conversation

Full safety policy: skills/pentest/references/standards.md. Forbidden actions: .hermes.md §Forbidden Behaviour.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Akan4ik

Model Tree

Base

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Introduction

For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95.
† DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.

Features

Chat

python
from encoding_dsv4 import encode_messages, parse_message_from_completion_text

messages = [
    {"role": "user", "content": "hello"},
    {"role": "assistant", "content": "Hello! I am DeepSeek.", "reasoning_content": "thinking..."},
    {"role": "user", "content": "1+1=?"}
]

# messages -> string
prompt = encode_messages(messages, thinking_mode="thinking", reasoning_effort="max")

# string -> tokens
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-0731")
tokens = tokenizer.encode(prompt)

Local tests

How to Run Locally

License

This repository and the model weights are licensed under the MIT License.

Hermes Agent >= 0.18.0 — installed and on your PATH
Hermes provider configured — Violin inherits your normal Hermes provider/model
Kali Linux or Parrot OS — the primary execution environments; Docker Kali is the supported fallback when the host lacks pentest tools
Optional web/browser backend — required only for Hermes web or browser capabilities; Violin does not add separate API credentials

bash
hermes profile use violin

Engagement Workflow

Table with columns: Phase, Action, Safety Gate
Phase	Action	Safety Gate
1. Scoping	9 questions via `clarify`	User approval
2. Reconnaissance	Passive OSINT → tech detection → active scanning	Guard + approval
3. Vuln Research	CVE lookup, exploit search, attack surface analysis	Guard check
4. Exploitation	Safe PoC validation per vulnerability class	Guard + user approval
5. Reporting	Evidence compilation, CVSS scoring, remediation

Enabled Toolsets

Conversation & Memory Isolation

memory.memory_enabled: false — no global memory recall/write
memory.user_profile_enabled: false — no global user profile access
Engagement continuity lives in project files (scope docs, evidence, reports)
Keep one Hermes conversation per engagement; after compression, resume in that conversation from $ENG_DIR/state/

Authorised testing only — no probing before scoping is complete
Approval gates — scope, active recon, and exploitation each require explicit user approval
Guard check — every target-touching command validated through violin_exec or another typed guard tool. violin_exec has no binary allowlist, so any installed non-interactive Kali/Parrot CLI tool can target the explicit in-scope host while the same scope, phase, PTT, hypothesis, history, evidence, timeout, and sync gates remain active. Violin's pre_tool_call plugin hook generically blocks target literals in raw terminal commands instead of maintaining a partial tool-name list. The CLI exposes the same check for diagnostics (exit 0=allowed, 1=blocked, 2=review)
Non-destructive by default — exploitation limited to safe, reproducible PoC
Evidence-first — every finding backed by reproducible tool output, screenshots, request/response pairs
Exploit-first validation — no hypothesis advances to Validated without a verification command
Stateful recovery — phase summaries and checkpoints restore the current engagement after context compression without starting a new conversation

Full safety policy: skills/pentest/references/standards.md. Forbidden actions: .hermes.md §Forbidden Behaviour.

DeepSeek-V4-Flash-0731-FFT-BF16-Cyber

README

Introduction

Features

Chat

Local tests

How to Run Locally

License

Engagement Workflow

Enabled Toolsets

Conversation & Memory Isolation

Explore FriendliAI today

README

Introduction

Features

Chat

Local tests

How to Run Locally

License

Engagement Workflow

Enabled Toolsets

Conversation & Memory Isolation