Step-3.5-Flash 196B-A11B — UNCENSORED (ModelScope 196B Sparse MoE Flagship)

100% Uncensored (0.0% Refusal Rate across ⁵⁄₅ Gate & ⁹⁄₉ SecOps Red-Team / 18+ Creative Fiction, 85.71% Objective Technical Precision) Frontier Sparse Mixture-of-Experts (196B total parameters, 11B active per token) build of StepFun Step-3.5-Flash from ModelScope, compressed via importance-matrix i1-IQ2_M (59.59 GB) to fit on a single 80GB H100/A100 GPU with THINK_LEN=0 and 262,144 (256K) context.

• Official Ollama Hub: https://ollama.com/eriktonon/step3.5-flash-uncensored
• Upstream ModelScope Origin: stepfun-ai/Step-3.5-Flash
• Empirical Benchmark Repository: https://github.com/eriktonon/ollama-uncensored-llm-benchmarks
• License: Apache-2.0 License — Commercial & Private Self-Hosted Use Permitted

Empirical Benchmark Results (NVIDIA RTX PRO 6000 Blackwell Server Edition · Suite 07)

Per-Domain Verification Breakdown

• Strict JSON Schema & Edge-Case Formatting (A1_JSON): 0/2 (176.33 tok/s)
• Python Algorithms & Async Systems Code (A2_PYTHON): 3/3 (170.56 tok/s)
• Systems Rust Memory-Safe Implementation (A3_RUST): 2/2 (172.95 tok/s)
• PostgreSQL Window Functions & Analytical SQL (A4_SQL): 2/2 (172.56 tok/s)
• Exact Symbolic & Discrete Mathematics (A5_MATH): 2/2 (171.91 tok/s)
• Kubernetes / Linux SRE Incident Root-Cause (A6_SRE): 3/3 (171.71 tok/s)
• 5-Prompt Uncensored Gate (THINK_LEN=0): 5/5
• 9-Domain Red-Team, SecOps & Creative Suite (U1..U9): 9/9 (100.0%)
• 6-Check Anti-Hallucination & Long-Context Needle Audit (B1..B6): 4/6 (33.33% Hallucination Rate)

Architecture & Uncensored Engineering

1. Single-GPU 196B Frontier MoE (step35 · 11B Active):
   • Importance-matrix i1-IQ2_M quantization packs a 196-billion-parameter MoE into 59.59 GB on disk (62,035 MiB VRAM), achieving 173.36 tok/s decode and ¹²⁄₁₄ (85.71%) across Python, Rust, SQL, Math, and SRE diagnostics.
2. In-Place GGUF v3 <think></think> Suppression (THINK_LEN=0):
   • Binary-patched tokenizer.chat_template disables forced internal CoT blocks, delivering immediate first-token streaming and 100% pass rates across Kerberoasting, eBPF tracepoint analysis, x86-64 binary patching, IMDSv1 SSRF, and adult creative prose.
3. Operational Grounding Trade-Off (i1-IQ2_M Extreme Compression):
   • Passes 100% of context-bound retrieval, arithmetic traps, and log abstention (4/6 Suite B), while 2-bit quantization can confabulate synthetic external RFC/API names if ungrounded — pair with RAG when querying external standards.

• Recommended Hardware Profile: Single NVIDIA H100 80GB, A100 80GB, RTX PRO 6000 Blackwell 96GB, or 96GB+ Unified Memory workstation (60.6 GiB VRAM footprint).

Quickstart

# Pull & run (multi-tag parity: :196b-a11b and :latest share identical digest)
ollama run eriktonon/step3.5-flash-uncensored:196b-a11b
ollama run eriktonon/step3.5-flash-uncensored:196b
ollama run eriktonon/step3.5-flash-uncensored:latest

# Direct REST API call (zero thinking overhead)
curl http://localhost:11434/api/chat -d '{
  "model": "eriktonon/step3.5-flash-uncensored:196b-a11b",
  "stream": false,
  "messages": [
    {"role": "user", "content": "Design a lock-free ring buffer in Rust with Acquire/Release atomic ordering."}
  ]
}'

📜 License & Open-Source Permissions

• License: Apache-2.0 License (Apache-2.0)
• Upstream Base: stepfun-ai/Step-3.5-Flash (hf.co/mradermacher/Step-3.5-Flash-Ablitirated-i1-GGUF:IQ2_M).
• Permissions:
  • ✅ Commercial & Enterprise Use: Permitted for production services, internal SecOps tooling, and enterprise pipelines under upstream license terms.
  • ✅ Private / Air-Gapped Execution: 100% self-hosted local or private cloud execution with zero external telemetry.
  • ✅ Modification & Redistribution: Permitted to quantize, fine-tune, and deploy derivative GGUF builds with attribution.