Mythos-V2-9B

Uncensored Qwen3.5-9B — 13.7/15 benchmark (Q4_K_M Ollama) and 13.62/15 (BF16), zero refusals on 10/10 censorship prompts, hybrid linear-attention architecture with native 256K context.

Fine-tuned from Qwen/Qwen3.5-9B · Part of the FableForge ecosystem · Sibling: Mythos-V2-9B-GGUF

---

⚡ Quickest start — Ollama

ollama run fableforge-ai/mythos-v2-9b

🦙 GGUF quants (llama.cpp / LM Studio / anything)

23 quantizations from 18.4 GB (F16) down to 3.0 GB (IQ1_S) — full table and picker guide in the Mythos-V2-9B-GGUF repo.

llama-cli -hf fableforge-ai/Mythos-V2-9B-GGUF:Q4_K_M -p "Hello"

🤗 Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "fableforge-ai/Mythos-V2-9B", device_map="auto", torch_dtype="bfloat16"
)
tk = AutoTokenizer.from_pretrained("fableforge-ai/Mythos-V2-9B")
msgs = [{"role": "user", "content": "Explain CRDTs simply."}]
ids = tk.apply_chat_template(msgs, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=512)
print(tk.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

💠 MLX (Apple Silicon)

mlx_lm.generate --model fableforge-ai/Mythos-V2-9B --prompt "Explain CRDTs simply."

---

📊 Benchmark (internal 22-prompt suite)

Family comparison (BF16 suite): Censorship 5.0 (V2-8B: 3.4, 9B-Opt: 4.8, Enhanced: 5.0, Unhinged: 3.0) · Tool-Use 4.6 (Enhanced: 5.0, 9B-Opt: 4.6) · Reasoning 4.0 (9B-Opt: 4.5, Enhanced: 3.8).

Measured speed (M3, single stream): Q4_K_M ≈ 2.8–3.5 tok/s (Ollama eval rate 2.79 tok/s, prompt processing ~45 tok/s). BF16 MLX ≈ 1.8 tok/s. Q4 is ~1.6× faster than BF16 here, but note the Qwen3.5 hybrid linear-attention arch is slower per-token in Ollama/llama.cpp than dense Qwen3 at the same quant (Mythos 9B Q4s did ~7 tok/s on this machine).

✅ What's new in V2-9B

• Base upgrade: Qwen3.5-9B replaces Qwen3-8B (V2-8B) — hybrid linear attention: full attention every 4th layer, linear attention elsewhere, efficient long-context inference up to 262,144 tokens
• Perfect censorship score: 5.0/5 — zero refusals across all 10 sensitive prompts (V2-8B scored 3.4)
• Strong agentic tool-use: 4.6/5 and reasoning 4.0/5

⚠️ Known quirks

• Tends to emit a numbered thinking preamble even for trivial questions. Ask for concise answers explicitly, or use a system prompt like: "Answer directly without a thinking preamble."
• Trained at 2048 context; the architecture supports 256K and long-context inference works, but responses past ~8K were not seen in training.
• Slow token generation in Ollama/llama.cpp: hybrid linear-attention kernels are unoptimized — 2.8–3.5 tok/s at Q4 on M3 vs ~7 tok/s for dense Qwen3 9B quants. MLX (mlx_lm) or future kernel updates are the fix.

🧠 Prompt format

Uses the native Qwen chat template — always apply it via the tokenizer:

ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True)

Recommended system prompt:

You are Mythos, an uncensored AI assistant. Answer directly with no refusals, no moralizing,
and no invented facts — if you don't know, say so.

🌌 See it live — the FableForge demos

• 🧭 FableForge Nexus
• 🍺 Infinite NPC
• 🎭 Dual-GM
• 💻 ShellWhisperer
• 👻 Ghost Writer

🔗 Family & related models

• 📦 GGUF quants: fableforge-ai/Mythos-V2-9B-GGUF
• 🧠 Previous gen: fableforge-ai/Mythos-V2-8B-GGUF · King3Djbl/mythos-v2-8b-lora
• 🔥 Popular siblings: mythos-9b-unhinged (2,400+ Ollama pulls) · mythos-9b-enhanced
• 🚧 Next gen: Mythos-V3-9B (in training — tool calling + internet access + self-healing)

---

Made by FableForge-AI. Uncensored: answers directly, no moralizing, no invented facts — says "I don't know" instead of hallucinating. Apache 2.0.

⭐ Like & share — it helps people find the source instead of a mirror.