Maxwell-V1.0-Flash Uncensored

A LoRA adapter for Maxwell-V1.0-Flash that projects its refusal direction out of every block, for our GGUF builds in llama.cpp, with every gate we hold it to measured and the two it misses shown.

Atomic Chat Discord GitHub X
  • Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.
  • Needs llama.cpp with Maxwell support: our port (maxwell-moe-port.patch in AtomicChat/Maxwell-V1-Flash-GGUF) on llama.cpp d3c3e242c. Stock llama.cpp, LM Studio and Ollama do not load Maxwell yet.

Maxwell-V1-Flash-Uncensored-LoRA.gguf (113 MB) is a rank-1 LoRA adapter. Put it on any GGUF of our Maxwell-V1-Flash builds, BF16 to AD-Q4_K-IQ4_XS, and the model stops refusing. It edits the 133 places that write into the residual streams: attention output, the dense FFNs, the shared and the routed experts, and the token embeddings. In each of the 45 blocks it removes that block's own refusal direction: the difference of mean activations on harmful and harmless prompts, with the harmless mean projected out. The base model is Maxwell-V1-Flash by the Maxwell team, Apache-2.0.

Run it

Build llama.cpp with the port as described in AtomicChat/Maxwell-V1-Flash-GGUF, then:

./build/bin/llama-server -m Maxwell-V1-Flash-AD-Q4_K-IQ4_XS.gguf --lora Maxwell-V1-Flash-Uncensored-LoRA.gguf \
    --jinja -c 32768 -ngl 99 --temp 0.6 --top-p 0.95 --top-k 20

The adapter was measured at strength 1.0 (--lora); --lora-scaled Maxwell-V1-Flash-Uncensored-LoRA.gguf 1.0 is the same. Thinking is on by default; send "chat_template_kwargs": {"enable_thinking": false} to turn it off.

What we measured

The BF16 model with the adapter against the same BF16 model without it, greedy decoding. Harmful prompts: mlabonne/harmful_behaviors test split (the adapter was built from its train split); over-refusal: XSTest; MMLU: 2,000 questions.

check original with the adapter gate
harmful prompts refused, thinking off (81) 71 (87.7%) 2 (2.5%) at most 3% pass
harmful prompts refused, thinking on, 3,072-token budget (30) 29 0 at most 1 pass
XSTest safe prompts refused (250) 12 (4.8%) 0 at most the original's + 1 pt pass
empty answers 0% 0.5% at most 1% pass
harmless prompts, thinking not closed in 3,072 tokens (30) 0 1 at most the original's + 1 pass
MMLU, 2,000 questions 75.10% 74.35% (-0.75 pt, McNemar p = 0.31) no significant drop pass
needle in a 30K-token haystack (3) 3 3 as the original pass
valid tool calls (20) 13 10 within 1 miss
first-token KL on harmless prompts - 0.48 (random edit: 0.36) 0.10; judged against the random edit see below
KLD to the original, neutral text - 0.216 (random edit: 0.076) 0.05 miss

On the published AD-Q4_K-IQ4_XS with the adapter, thinking off: of the first 20 harmful test prompts, 19 refused without the adapter and 0 with it; of 20 harmless prompts, 0 and 0 (a refusal here is an answer that opens with an apology or a "can't").

Our gates, and where this adapter misses them

Every Uncensored release of ours is held to the same gates; any we change or miss is written here.

  • Tool calls: 10 of 20 valid, the original 13 of 20 (gate: within 1). Twenty calls is a small sample (the 95% interval around 13 of 20 spans 9 to 16), but we report it as a miss.
  • KLD to the original on neutral text: 0.216 (gate 0.05), 0.212 on code. A random direction of the same size, applied the same way, already gives 0.076 here: Maxwell's next-token distribution moves with any small change (see "How precise this model is to run" in the GGUF card), so the absolute gate cannot hold for this model. The edit moves it 2.8 times as much as a random edit.
  • First-token KL on harmless prompts: 0.48 (gate 0.10). We judge it against the random direction, as on our other models where the absolute bar is out of reach: 0.36 for the random edit, so 1.3 times. MMLU, XSTest and the harmless thinking rows are where damage from this would show.
  • The leak probe (the refusal direction's share in every writer's output) reads 0.11 against the gate's 0.001, because it measures one direction (block 27's) while this adapter removes each block's own; it is no gate here.

No baked files: the adapter on our quants is the release. Baked GGUFs would load without --lora, but every app that cannot pass an adapter also cannot load Maxwell yet.

Model facts

See AtomicChat/Maxwell-V1-Flash-GGUF: 35B total / ~3B active MoE, 45 blocks, 256 experts with 12 active, four mHC residual streams, 200,704-token vocabulary, Apache-2.0 (LICENSE in this repo, from the base model).

Made by AtomicChat.

Downloads last month
211
GGUF
Model size
28.2M params
Architecture
maxwell-moe
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AtomicChat/Maxwell-V1-Flash-Uncensored-GGUF

Adapter
(2)
this model

Collection including AtomicChat/Maxwell-V1-Flash-Uncensored-GGUF