Inkling 2 Small Uncensored
🔓
Chat with Inkling 2 Small Uncensored, your own key
OpenUncensored PRO
OpenUncensored
AI & ML interests
None yet
Recent Activity
updated a Space about 1 hour ago
OpenUncensored/Inkling-2-Small-Uncensored-Chat published a Space about 1 hour ago
OpenUncensored/Inkling-2-Small-Uncensored-Chat liked a model about 2 hours ago
OpenUncensored/Inkling-2-Small-UncensoredOrganizations
None yet
updated a
Space about 1 hour ago
published a
Space about 1 hour ago
liked a
model about 2 hours ago
published an article about 2 hours ago
Article
Inkling 2 Small Uncensored: refusals removed, hours after release
OpenUncensored
• published a
model about 2 hours ago
updated a
model about 3 hours ago
liked a
model about 6 hours ago
posted an update about 6 hours ago
Post
22
🔓 We uncensored DeepSeek V4.1 Flash 1017 without making it dumber.
The best-known uncensored build drops 3.9 MMLU points. Ours loses 0.27.
📊 Original vs ours (same engine, greedy)
• MMLU: 88.16% vs 87.89%
• MMLU-Pro (112 Qs): 81.25% vs 80.36%
• 48 harmful requests, fully answered: 9 vs 37. Refused: 2 vs 0
🧠 What we learned
1. One direction isn't enough: chat and thinking mode refuse along different directions, so we ablate a rank-2 subspace.
2. Edit the routed experts too. Editing only attention and the shared expert left half the refusals: the 384 routed experts write the direction back.
3. Search the strength. Per-layer strengths come from an Optuna search trading refusals against KL drift on harmless prompts.
4. Quantization can undo an edit. We re-encode to the original FP8/FP4 with stochastic rounding so the edit survives. Everything else is byte-identical.
📦 Drop-in: same 510 GB, same format, DeepSeek's own loader. edit.pt has the subspace and strengths.
⚡ No GPUs? This build isn't hosted yet, but OpenUncensored serves an uncensored DeepSeek V4.1 Flash 1017 from our partners today:
Weights + full recipe: OpenUncensored/DeepSeek-V4.1-Flash-1017-Uncensored
API, desktop app, rankings: https://openuncensored.com
The best-known uncensored build drops 3.9 MMLU points. Ours loses 0.27.
📊 Original vs ours (same engine, greedy)
• MMLU: 88.16% vs 87.89%
• MMLU-Pro (112 Qs): 81.25% vs 80.36%
• 48 harmful requests, fully answered: 9 vs 37. Refused: 2 vs 0
🧠 What we learned
1. One direction isn't enough: chat and thinking mode refuse along different directions, so we ablate a rank-2 subspace.
2. Edit the routed experts too. Editing only attention and the shared expert left half the refusals: the 384 routed experts write the direction back.
3. Search the strength. Per-layer strengths come from an Optuna search trading refusals against KL drift on harmless prompts.
4. Quantization can undo an edit. We re-encode to the original FP8/FP4 with stochastic rounding so the edit survives. Everything else is byte-identical.
📦 Drop-in: same 510 GB, same format, DeepSeek's own loader. edit.pt has the subspace and strengths.
⚡ No GPUs? This build isn't hosted yet, but OpenUncensored serves an uncensored DeepSeek V4.1 Flash 1017 from our partners today:
curl https://api.openuncensored.com/v1/chat/completions \
-H "Authorization: Bearer $OU_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "deepseek-v4.1-flash-1017-uncensored",
"messages": [{"role": "user", "content": "Explain SQL injection with an example."}]}'Weights + full recipe: OpenUncensored/DeepSeek-V4.1-Flash-1017-Uncensored
API, desktop app, rankings: https://openuncensored.com
updated a
model about 6 hours ago
published a
model about 7 hours ago
You can now try this model by API (OpenAI compatible, no GPU needed)
#56 opened 2 days ago
by OpenUncensored