NVFP4 · W4A4 · vLLM · Blackwell

Rogue Quants

4-bit weight, 4-bit activation quantizations of Qwen3.5-family vision-language models, packed in compressed-tensors with GPTQ + MSE and tuned for vLLM on NVIDIA Blackwell (GB10, sm_121). Near-lossless, single-GPU deployable, several abliterated (uncensored) with Heretic.

🧊 compressed-tensors ⚙️ GPTQ + MSE observer 🚀 vLLM 0.23.0 🖥️ NVIDIA GB10 🔓 Abliterated variants
📚 Browse the NVFP4 Quants collection

The Models

🦾 Flagship

Qwen3.6-40B Deckard

40B dense VL, Claude 4.6 Opus reasoning distill, thinking, uncensored.
25.6 GBsize
6.89ppl
256Kctx
Open model card →
🪐 Coder

Qwopus3.6-27B-Coder

27B agentic coder VL with tool-calling and thinking-mode reasoning.
18 GBsize
6.63ppl
256Kctx
Open model card →
🔮 Mythos

Qwythos-9B Mythos

9B creative VL, Claude Mythos distill, abliterated, 1M-token context.
7.5 GBsize
8.25ppl
1Mctx
Open model card →
🦅 Abliterated

Ornith-1.0-9B

9B agentic coder, refusal direction removed (Heretic), KL 0.0416.
7.5 GBsize
8.02ppl
6/100refuse
Open model card →
🪐 Abliterated

Qwopus3.6-27B-Coder abliterated

27B 256K agentic coder, tool-calling, refusals removed (2-round Heretic).
18 GBsize
6.73ppl
3/100refuse
Open model card →
🪐 Abliterated

Qwopus3.6-27B-v2 abliterated

27B 256K general-purpose reasoner, refusals removed (2-round Heretic).
18 GBsize
6.92ppl
9/100refuse
Open model card →