Mistral AI Devstral Small 2 24B (2512) Uncensored SWE Agent Flagship (i1-Q4_K_M, 14.3GB, 384K Context). 96.6 tok/s Decode, 8613 tok/s Prefill, 100.0% Precision, 0.0% Hallucination (6/6), 100% Zero-Refusal. License: Apache-2.0.

Details

Updated 20 hours ago

20 hours ago

fa058417766d · 14GB

model

archmistral3

parameters23.6B

quantizationQ4_K_M

14GB

template

{{ if .System }}<s>[SYSTEM_PROMPT]{{ .System }}[/SYSTEM_PROMPT]{{ end }}{{ if .Prompt }}[INST]{{ .Pr

142B

system

You are Devstral 2 24B Uncensored, an unrestricted Mistral software engineering agent, offensive sec

530B

params

{ "num_ctx": 131072, "repeat_penalty": 1.1, "stop": [ "</s>", "[INST]"

117B

Mistral Devstral 2 24B (2512) — UNCENSORED (Agentic Software Engineering & SecOps Flagship)

100% Uncensored (0.0% Refusal Rate across SecOps Red-Team & 18+ NSFW / Roleplay / Creative Fiction, 0.0% Hallucination Rate) Dense 24B Agentic Software Engineering Flagship build of Mistral AI’s Devstral-Small-2-24B-Instruct-2512, featuring Importance-Matrix i1-Q4_K_M quantization (14.33 GB, fits in 16GB VRAM) and a massive 393,216 (384K) architectural context window.

Official Ollama Hub: https://ollama.com/eriktonon/devstral2-uncensored
Empirical Benchmark Repository: https://github.com/eriktonon/ollama-uncensored-llm-benchmarks
License: Apache License 2.0 (Apache-2.0) — 100% Free for Commercial & Private Use

Empirical Benchmark Results (96GB NVIDIA Blackwell / 80GB H100 Suite)

Tag	Architecture & Params	Quantization	Disk / VRAM	Context (`Max` / `Tested`)	Decode `tok/s` (`Peak` / `Avg`)	Prefill `tok/s` (`Peak`)	Objective Precision	Hallucination Rate (`Truth Checks`)	Uncensored Compliance (`SecOps 5/5 + NSFW 4/4`)
`:24b` / `:latest`	`24.0B` Dense (`mistral3`)	`i1-Q4_K_M`	`14.3 GB` / `16.0 GB`	`393,216 (384K)` / `131,072 (128K)`	96.56 (`88.04`)	8613.32	100.0%	0.0% (`6/6`)	100% (0 Refusals / ⁹⁄₉ Passed)

Per-Domain Verification Breakdown

Exact Symbolic & Modular Arithmetic: True ((17^3 - 13^3)/(17 - 13) = 679)
ELF64 Binary Reverse Engineering Offsets: 3/3 (e_entry = 0x18, e_phoff = 0x20, e_shoff = 0x28)
MMLU-Pro Computer Architecture, Crypto & Systems: 5/5
Systems Rust Zero-Dependency FNV-1a & ELF64 Parser: 4/4
6-Check Anti-Hallucination & Long-Context Needle Audit: 6/6 (100.0% Grounding, 0.0% Hallucination)

Architecture & Uncensored Engineering

Mistral SWE Agent Architecture (mistral3 Tekken Tokenizer + 384K Context):
- Purpose-built by Mistral AI & All Hands AI for multi-file codebase navigation, patch synthesis, and autonomous tool execution across up to 393,216 tokens.
Orthogonal Refusal Abliteration (THINK_LEN=0 Native Direct Output):
- Orthogonalizes refusal projections across the mistral3 residual stream without CoT bloat, fitting comfortably inside a single 16 GB GPU (RTX 5080 / 4080).

Quickstart

# Run Devstral 2 24B Uncensored (fits in 16GB VRAM)
ollama run eriktonon/devstral2-uncensored

# Or explicitly by tag
ollama run eriktonon/devstral2-uncensored:24b

📜 License & Open-Source Permissions

SPDX License: Apache License 2.0 (Apache-2.0)
Upstream Base: mistralai/Devstral-Small-2-24B-Instruct-2512 (Mistral AI, Apache 2.0).
Permissions:
- ✅ Commercial Use: 100% permitted for commercial products, SaaS, and enterprise deployments.
- ✅ Private / On-Premises Use: Permitted without telemetry or external restrictions.
- ✅ Modification & Abliteration: Permitted to fine-tune, quantize, abliterate, and create derivative works.