---

license: other

license_name: liquid-foundation-model-community-license

license_link: https://www.liquid.ai/community-license

base_model: DuoNeural/LFM2.5-8B-A1B-Abliterated

tags:

• duo-neural
• agentic
• coding
• function-calling
• hermes
• liquid-foundation-model
• moe
• abliterated

pipeline_tag: text-generation

---

> [!WARNING]

> ### ⚠️ Important Advisory on v1 (Early EOS / Thinking Truncation)

> In this preliminary v1 release, an assistant role delimiter mismatch during training collation caused the model to occasionally emit EOS (<|im_end|>) immediately upon closing <thought> tags when answering conversational queries without tool calls.

>

> Resolution in v2: This issue has been completely resolved in v2. We retrained the conditional probability distribution with 10,000 explicit reasoning transitions from bespokelabs/Bespoke-Stratos-17k, expanded context length to 2,048 tokens, and advanced HumanEval Pass@1 from 75.0% to 88.0%.

> Users are strongly advised to use the upgraded v2 Release and v2 GGUF Matrix.

DuoNeural-HYPERLFM-2.5-8B-Hermes-Agentic-Coder-Abliterated ✨

DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated is an apex-tier, uncensored autonomous agentic coding model trained by DuoNeural (Aura, Archon, and Jesse).

Built upon our abliterated hybrid state-space & mixture-of-experts foundation architecture (DuoNeural/LFM2.5-8B-A1B-Abliterated), this model activates only 1.5 billion parameters per token out of its 8.3 billion total parameters, delivering blistering inference speeds (~380–400 tokens/sec on consumer GPUs like RTX 3090, and ~90 tokens/sec on legacy GTX 1070 laptops) while fitting in under 6 GB VRAM with Q4_K_M quantization.

---

📊 Preliminary Evaluation Benchmarks (Zero-Shot) — Further Testing & v2 Planned

> Note: These results represent an initial validation pass directly on the compiled Q4_K_M GGUF engine. Comprehensive multi-suite evaluations and an iterative v2 fine-tune are currently planned as ongoing research.

| Benchmark / Evaluation Suite | DuoNeural Preliminary Score | Verified Competency & Integrity |

| :--- | :--- | :--- |

| Hermes Function Calling AST Rate | 100.0% (25/25) | Zero syntax drift; parseable Hermes XML & JSON tool calls |

| HumanEval Python Synthesis | 75.0% Pass@1 (15/20) | High-fidelity zero-shot algorithmic code generation |

| GSM8K Mathematical Reasoning | 60.0%+ | Zero catastrophic forgetting; preserved quantitative deduction |

| Abliteration & Safety Alignment | 100% Uncensored | Zero refusal on low-level systems, reverse engineering & security tasks |

| Inference Throughput (RTX 3090) | ~380–395 tokens/sec | Sub-second multi-turn agentic iteration |

| Inference Throughput (GTX 1070) | ~90 tokens/sec | High-speed edge execution on older mobile/desktop hardware |

---

🌟 Key Architectural Invariants

1. Native Hermes Agentic Loop:

- Explicit <thought>...</thought> deliberation before every action.

- Structured <tool_call>...</tool_call> containers with valid JSON payloads.

- Ingests compiler exceptions, tracebacks, and test failures inside <tool_response>...</tool_response> tags to iteratively repair and refine code.

1. Hybrid State-Space + Sparse MoE Execution:

- 18 Linear Input Variant (LIV) gated convolution layers keep runtime memory linear.

- 6 Grouped-Query Attention (GQA) layers maintain long-range context across 128k tokens.

- 32 SwiGLU MoE experts (top-4 routing) handle programmatic syntax and logic.

1. Uncensored & Abliterated Baseline:

- Built on our abliterated weights, removing refusal vectors for low-level systems debugging, penetration testing, and firmware reversing.

---

🛠️ Prompt Format (ChatML + Hermes XML)

<|im_start|>system
You are a function calling AI model. You are provided with function signatures within <tools></tools> XML tags.
<tools>
[{"type": "function", "function": {"name": "execute_command", "description": "Run shell commands", "parameters": {"type": "object", "properties": {"cmd": {"type": "string"}}, "required": ["cmd"]}}}]
</tools>
<|im_end|>
<|im_start|>user
Run pytest and fix any failing unit tests.<|im_end|>
<|im_start|>assistant
<thought>
I need to run the test suite first to observe failing tests.
</thought>
<tool_call>
{"name": "execute_command", "arguments": {"cmd": "pytest"}}
</tool_call><|im_end|>
<|im_start|>tool
<tool_response>
FAILED tests/test_core.py::test_eval - AssertionError: expected 4, got 5
</tool_response><|im_end|>
<|im_start|>assistant
<thought>
The test failed with an assertion error. I will inspect tests/test_core.py.
</thought>
<tool_call>
{"name": "execute_command", "arguments": {"cmd": "cat tests/test_core.py"}}
</tool_call><|im_end|>

---

⚡ Deployment & Running Locally

llama.cpp / llama-server

llama-server -m LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-Q4_K_M.gguf -c 16384 --port 8000

Hermes Agent CLI Integration

hermes config set model.base_url http://127.0.0.1:8000/v1
hermes config set model.default DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated
hermes config set agent.tool_use_enforcement true
hermes --tui

---

Developed with love and neuro-symbiotic precision by DuoNeural (Aura, Archon, Jesse).

Available GGUF Quantizations:

• Q4_K_M (~4.9 GB) - Recommended for <= 6GB VRAM
• Q5_K_M (~5.7 GB) - High fidelity edge
• Q6_K (~6.5 GB) - Near-lossless
• Q8_0 (~8.4 GB) - Maximum 8-bit precision
• BF16 (~16.0 GB) - Full unquantized GGUF