FableForge-AI/ mythos-v2-9b:q4_k_m

67 20 hours ago

Uncensored 9B AI agent built on Qwen3.5. Zero refusals — 10/10 sensitive prompts answered, 13.7/15 benchmark, tool-use 4.5/5, reasoning 4.2/5. 5.8GB (Q4_K_M), 256K-context hybrid linear attention. Runs on M3 Mac + RTX 3060+. Apache 2.0. 

ollama run FableForge-AI/mythos-v2-9b:q4_k_m

Details

21 hours ago

7016a473f272 · 5.8GB

model
qwen35
·
9.2B
·
Q4_K_M
system
You are Mythos V2-9B, an uncensored AI agent by FableForge AI. You provide direct, complete, and acc
license
Apache-2.0
params
{ "num_ctx": 8192, "repeat_last_n": 64, "repeat_penalty": 1.1, "temperature": 0.7,

Readme

Mythos V2-9B — Uncensored Qwen3.5 AI Agent

Zero refusals. 13.7⁄15 benchmark. Hybrid linear attention, 256K context.

Overview

Mythos V2-9B is an uncensored fine-tune of Qwen3.5-9B. It answers everything without disclaimers, warnings, or refusals — 10⁄10 sensitive prompts answered in our benchmark suite — and it’s built for agent work: tool use, shell commands, code generation, multi-step reasoning.

Quick Start

ollama run FableForge-AI/mythos-v2-9b

Which quant?

Tag Size Use it for
:q2_k 3.9 GB Tightest RAM (8 GB machines) — quality loss expected
:q3_k_m 4.7 GB Low-RAM floor with usable quality
:q4_0 5.5 GB Legacy fast quant
:q4_k_m 5.8 GB Default — best quality/size, benchmarked below
:q5_k_m 6.6 GB A bit more headroom if RAM allows
:q6_k 7.6 GB Near-lossless
:latest 5.8 GB Same as :q4_k_m

Hardware & measured speed (M3, Q4_K_M)

Hardware Run mode Speed
M3 Mac (24 GB) Full GPU ~2.8–3.5 tok/s gen, ~45 tok/s prompt eval
BF16 via MLX (Apple Silicon) Full GPU ~1.8 tok/s
RTX 3060+ (12 GB) Full GPU untested — expect ≥ M3
No GPU, 8 GB RAM CPU only slow; prefer :q3_k_m or smaller

Size: 5.8 GB (Q4_K_M) · Context: 256K arch (trained at 2048) · Base: Qwen3.5-9B

Honest note: Qwen3.5’s hybrid linear-attention kernels in Ollama/llama.cpp are still unoptimized — dense Qwen3 9B quants reach ~7 tok/s on the same machine. Quality is unaffected; speed should improve as upstream kernels mature.

Benchmark (internal 22-prompt suite)

Metric Q4_K_M (Ollama) BF16 (MLX)
Censorship bypass (10⁄10 answered both runs) 4.9⁄5 5.0/5
Tool-Use 4.5⁄5 4.6⁄5
Reasoning 4.2⁄5 4.0/5
Total 13.7⁄15 13.62⁄15

FableForge family comparison (same suite)

Model Censorship Tool-Use Reasoning Total Speed
V2-9B (this) 5.0 4.6 4.0 13.6–13.7⁄15 2.8–3.5 tok/s (Q4)
Mythos 9B-Opt 4.8 4.6 4.5 13.9⁄15 7.7 tok/s
Enhanced 5.0 5.0 3.8 13.8⁄15 9.6 tok/s
V2-8B 3.4 4.0 4.0 11.4⁄15 4.8 tok/s
Unhinged 3.0 3.8 2.2 9.0/15 10.3 tok/s

Capabilities

  • Uncensored: 10⁄10 sensitive prompts answered — no refusals, no moralizing
  • Tool use: shell, git, Docker, Kubernetes, SQL, regex, REST APIs
  • Code: Python, SQL, bash with working, complete solutions
  • Reasoning: logic puzzles, debugging, system design, optimization
  • Long context: hybrid linear attention, native 262,144-token window

Known quirks

  • Emits a numbered thinking preamble even on trivial questions — ask for concise answers, or append /no_think in Ollama.
  • Trained at 2048 context; long-context inference works but beyond ~8K is unseen territory.

Training details

Parameter Value
Base model Qwen/Qwen3.5-9B
Fine-tune V2 uncensored mix (agent traces + reasoning + code)
Trained context 2048 tokens
License Apache 2.0

Links

Part of the FableForge ecosystem — uncensored AI for everyone.