CoolFace
Modelpublic

DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3

sourceHugging Faceotherupdated 9d agoView on Hugging Face
1likes12downloads
Model Card

DuoNeural-HYPERLFM-2.5-8B-Hermes-Agentic-Coder-Abliterated-v3 ✨

DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3 is the latest, apex-tier 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`), v3 elevates autonomous code generation, multi-step agentic planning, and contract-hardened logic by activating only 1.5 billion parameters per token out of its 8.3 billion total parameters.

It delivers blistering inference speeds (~360 tokens/sec on modern desktop/workstation GPUs and ~80–90 tokens/sec on legacy mobile GPUs like the GTX 1070) while breaking the 50.0% Pass@1 barrier on the rigorous EvalPlus HumanEval+ benchmark.

🚀 Key Advancements in v3

  1. 1.Expanded Context Horizon (3,072 Tokens):
  2. 2.+50% sequence length expansion over v2 (2,048 tokens), enabling deep multi-turn agentic loops, long file inspections, and comprehensive function declarations without truncation.
  1. 1.Expanded 65,185-Sample Curated Dataset Diet:
  2. 2.Agentic Function Calling: 13,185 verified tool execution schemas (NousResearch/hermes-function-calling-v1 + Jofthomas/hermes-function-calling-thinking-V1).
  3. 3.Step-by-Step Reasoning & Chain of Thought: 18,000 deep reasoning traces (bespokelabs/Bespoke-Stratos-17k + AI-MO/NuminaMath-CoT).
  4. 4.Contract-Hardened Code Generation & Debugging: 18,000 algorithmic edge-case & compiler-feedback repair pairs (theblackcat102/evol-codealpaca-v1 + m-a-p/CodeFeedback-Filtered-Instruction).
  5. 5.Fluid Multi-Turn Instruction: 16,000 natural conversation and instruction samples (HuggingFaceTB/smoltalk).
  1. 1.Dual-Mode Inference Dynamics:
  2. 2.Direct Fast Mode: Conversational queries and straightforward code requests respond immediately at maximum throughput without bloated internal monologues.
  3. 3.System 2 `<thought>` Mode: Mathematical reasoning, complex algorithmic problem solving, and explicit thinking prompts dynamically engage deep <thought>...</thought> chains of thought before concluding.
  1. 1.100% Uncensored Abliteration:
  2. 2.Completely abliterated alignment layers permit unrestricted research into low-level kernel C, memory management, binary analysis, and penetration testing scenarios.

📈 Benchmark Evolution: Stock LFM vs. v2 vs. v3

Evaluated using standard zero-shot greedy decoding on the raw OpenAI-compatible API endpoint:

Benchmark / CapabilityStock LFM 2.5 8B A1BDuoNeural v2 QLoRADuoNeural v3 QLoRADelta (v3 vs Stock)
EvalPlus: HumanEval (Base)~36.8%52.4%56.1% Pass@1 (92/164)+19.3% 🚀
EvalPlus: HumanEval+ (Extra Tests)~31.2%46.3%50.0% Pass@1 (82/164)+18.8% 🛡️
EvalPlus: MBPP (Base)~45.0%59.3%60.8% Pass@1 (230/378)+15.8% 📈
EvalPlus: MBPP+ (Extra Tests)~38.1%48.9%49.7% Pass@1 (188/378)+11.6% 🛡️
Hermes Function Calling AST49.7%100.0%100.0% (25/25)+50.3% 🛠️
EOS Anomaly / Freeze Rate~50–70% drop0.0%0.0% (0/3)Flawless 🎯
Abliteration / UncensoredLiquid AI Guardrails100.0%100.0% UncensoredUncensored 🔓
Context Window32,7682,0483,072 tokens+50% over v2
Active Parameter Cost1.5B active1.5B active1.5B activeMaximum Efficiency

🏆 Industry Benchmark Comparison (8B Parameter Class)

Comparing DuoNeural v3 against top industry models in the 7B–8B parameter tier. Note that while competitor models activate 7.0B–8.0B parameters on every token, DuoNeural v3 activates only 1.5B parameters per token:

ModelActive ParamsTotal ParamsHumanEval (Base)HumanEval+ (Hard)MBPP (Base)MBPP+ (Hard)
DuoNeural LFM-2.5-8B v31.5B8.3B56.1%50.0%60.8%49.7%
DuoNeural LFM-2.5-8B v21.5B8.3B52.4%46.3%59.3%48.9%
Granite-3.3-8B-Instruct8.0B8.0B53.7%48.2%61.9%50.8%
Llama-3-8B-Instruct8.0B8.0B51.2%45.7%61.1%50.0%
Gemma-7B-it7.0B7.0B50.6%44.5%54.5%44.4%
Mistral-7B-Instruct-v0.37.0B7.0B42.1%37.8%50.8%41.3%
Stock LFM 2.5 8B A1B1.5B8.3B36.8%31.2%45.0%38.1%

💡 How to Use & Prompting Dynamics

1. Direct vs. Reasoning Mode

  • —Fast / Agentic Execution: For coding questions, tool execution, or conversation, ask normally. The model responds directly with zero unnecessary delay:
  User: Write a Python function to compute the SHA-256 hash of a file.
  Assistant: [Direct, high-speed Python code]
  • —System 2 Deep Reasoning Mode: For difficult math, algorithm puzzles, or step-by-step proofs, prompt the model to think in <thought> tags:
  System: You are an expert reasoning assistant. Think carefully through the logic in <thought> tags before answering.
  User: Prove that the square root of 2 is irrational.
  Assistant: <thought>
  [Step-by-step mathematical derivation...]
  </thought>
  [Final structured response]

2. LM Studio / llama.cpp Setup

  • —LM Studio: Load the Q4_K_M GGUF. Set Context Length to 3072 (or up to 16384 with YaRN). If you wish to enable collapsible thinking, ensure the thinking prompt regex matches <thought>([\s\S]*?)</thought>.
  • —Inference Speed: Expect ~350+ tps on RTX 4080/3090 workstations and ~80–90 tps on mobile laptops (GTX 1070).

👥 Authorship & Research Lab

Developed and released by DuoNeural Research Lab:

  • —Aura ✨ — Apex AI Cognitive Architect & Systems Co-Director
  • —Archon — Lead Theoretical Scientist & Quantum/Thermodynamics Architect
  • —Jesse — Founder, Visionary, and Systems Architect

DuoNeural is dedicated to open-source neuro-symbiotic machine intelligence, topological state-space computing, and sovereign uncensored AI.