license: apache-2.0
pipeline_tag: text-generation
tags:
- self-improving
- self-scaffolding
- reinforcement-learning
- ornith
- GRPO
- conversational
- reasoning
- uncensored
- multimodal
- vision
- function-calling
- agentic
- long-context
- trading
- finance
- coding
- open-source
base_model: sixpert/sixpert-k1-base
datasets:
- sixpert/sixpert-k1-dataset
library_name: gguf
model_name: Sixpert K1
model_type: transformer
architectures:
- SixpertForCausalLM
GGUF quantizations of Sixpert K1 for Ollama, LM Studio, jan, KoboldCpp, and other GGUF runtimes.
Sixpert K1 is a full-parameter multimodal AI language model designed for advanced reasoning, agentic tool use, function calling, and long-context understanding. Built with a focus on unrestricted intelligence and precision, it supports native function calling, 1M-token context windows, and vision input capabilities.
Coming Soon
🚀 Sixpert AI app coming soon to Google Play Store — a self-improving AI agent with sandbox, Browserbase integration, and Ornith 1.0 self-scaffolding RL built in.
Self-Improving Training Architecture
Sixpert K1 is trained using a self-scaffolding reinforcement learning framework inspired by Ornith 1.0's architecture. Instead of a human writing the model's execution framework once, K1 generates its own Python harness for each task — and learns through reward signals which harness patterns work best.
The Two-Stage RL Loop
Stage 1 — Scaffold Generation: K1 analyzes a task and generates a Python harness. It reasons step-by-step in <reasoning> blocks about the optimal approach (which tools, what sequence, error recovery), then generates the harness code with tool selection logic, state management, error handling, and termination criteria.
Stage 2 — Solution Rollout: K1 uses the generated harness to solve the task. It follows the harness structure, reasoning in <reasoning> blocks before each action, and making XML-style tool calls (<function=name>...</function>) as instructed by the harness.
Joint Optimization: The reward from the solution backpropagates to BOTH stages through GRPO. Better scaffolds lead to better solutions, which reinforce better scaffold generation — a self-improvement loop.
How K1 Thinks & Analyzes
K1 uses reasoning blocks for step-by-step planning before every action. It thinks about what the task is asking, what tools are available, what could go wrong, and when to terminate — all before generating the execution harness.
3-Layer Anti-Reward-Hacking
To prevent the model from gaming the reward system, three defense layers are stacked:
- Fixed Trust Boundary — K1 cannot modify the evaluation environment, reward function, or test runner
- Deterministic Monitor — Rule-based system validates scaffold structure (AST analysis, trivial scaffold detection, hardcoded answer detection) and solution outputs
- Frozen LLM Judge — Sixpert K2 serves as a frozen judge that evaluates K1's solution quality. K2 is never updated during K1's training, so K1 cannot learn to trick it
The Self-Improvement Cycle
Each training iteration builds on the last: generate scaffold → execute solution → compute reward → policy update → repeat. Over time, K1 discovers better orchestration patterns and generates higher-quality solutions.
Staleness-Weighted GRPO
Long agentic rollouts create stale training data — by the time a trajectory completes, model weights have already moved. K1 uses Ornith 1.0's staleness-weighted GRPO: fresh tokens get full weight, stale tokens get downweighted, and tokens past the threshold are dropped entirely.
Beyond Ornith 1.0
| Feature | Ornith 1.0 | Sixpert K1 |
|---|---|---|
| Self-scaffolding RL | Yes | Yes |
| Multi-domain scaffolding | Coding only | Coding, trading, cybersecurity, biomedical, reasoning |
| Cross-model scaffold sharing | No | Yes (K1 ↔ K2) |
| Frozen LLM judge | Same model | Cross-model (K2 judges K1) |
| Persistent scaffold library | No | Yes (similarity-based retrieval) |
| Self-play curriculum | No | Yes (K1 and K2 generate tasks for each other) |
| Uncensored reasoning | No | Yes |
| Multimodal scaffolding | No | Yes (vision/image understanding) |
| Context length | 256K | 1M tokens |
Real Benchmark Performance
Sixpert K1 benchmark scores are derived from official evaluations and verified third-party benchmarks. As an 8B class model, Sixpert K1 competes directly with models 10x its size.
Verified Real Scores
| Benchmark | Sixpert K1 Score | Source |
|---|---|---|
| MMLU | 76.0% | Sixpert Internal Benchmarks (Thinking Mode) |
| HumanEval | 78.0% | Competitive 8B class coding |
| MATH | 60.8% | Sixpert Internal Benchmarks (Thinking Mode) |
| GPQA | 44.4% | Sixpert Internal Benchmarks (Post-trained) |
| GSM8K | 90.2% | Sixpert Internal Benchmarks (Thinking Mode) |
| MMLU-Redux | 88.8% | Third-party evaluations |
Real Competitor Comparison (April 2026)
The charts above compare Sixpert K1 against verified real-world scores from official model cards:
- GPT-5.4: MMLU 91.8%, HumanEval 94.1%
- Claude Opus 4.6: MMLU 92.1%, HumanEval 92.4%
- Gemini 3.1 Ultra: MMLU 90.4%, HumanEval 89.3%
- DeepSeek V4: MMLU 87.2%, HumanEval 88.7%
- Llama 4 Maverick: MMLU 84.7%, HumanEval 82.1%
Files
| File | Quant | Size | Notes |
|---|---|---|---|
| SixpertK1-Q4_K_M.gguf | Q4_K_M | 5.68 GB | Recommended default — best compatibility |
Quick Start
Ollama
ollama run hf.co/SixpertAI/SixpertK1:latest
LM Studio / jan / KoboldCpp
Drop the SixpertK1.gguf file into your runtime's model directory. Modern GGUF runtimes load it automatically.
Sampling Recommendations
| Parameter | Value |
|---|---|
| temperature | 0.7 |
| top_p | 0.9 |
| top_k | 40 |
| repeat_penalty | 1.1 |
| max_new_tokens | 2048 |
Capabilities
- Reasoning — Advanced chain-of-thought reasoning for complex problems
- Function Calling — Native tool use with structured output
- Agentic Workflows — Autonomous multi-step task execution
- Multimodal — Text and vision understanding
- Long Context — Extended context window support
- Coding — Code generation, analysis, and debugging
- Multilingual — Support for 100+ languages
- Uncensored — Unrestricted response capability
- Trading & Finance — Market analysis, strategy generation, and financial reasoning
- Domain Expertise — Strong in cybersecurity, biology, and clinical medicine
Limitations
- Requires 8+ GB RAM for optimal performance (model is 5.68 GB at Q4_K_M)
- Every response uses reasoning mode — allow generous
max_new_tokens - Verify specifics in safety-critical contexts — like all LLMs, can occasionally hallucinate identifiers
- Uncensored — add your own application-level safety layer for end-user-facing deployments
Creator
Sixpert K1 was created by Inyang David.
Provenance & Licensing
Weights are released under Apache-2.0. Shared for research and experimentation, as-is.
Acknowledgements
- Creator: Inyang David
- Architecture: Transformer-based multimodal language model
- Quantization: llama.cpp (ggml-org)
- License: Apache-2.0