Update, 21 November 2026: the files in this repository are now the uncensored model. They are built from Thinking Machines' original weights with AtomicChat's uncensoring edit merged in, and they replace the earlier censored files. A 6-bit file was added. The repository and the files were renamed to Inkling-2-Mini-Uncensored (old links redirect here). The model now answers requests the original refuses. The earlier files are still in this repository's commit history.
Hardware-aware GGUF quants of Inkling 2 Mini by Thinking Machines, built with the ROCmFPX weight formats for AMD GPUs (Vulkan and ROCm).
Inkling 2 Mini: 26.1B parameters, 24 layers, 128 experts per layer with 8 active per token (about 3.1B active parameters per layer pass). It is a looped model: layers 6 to 17 run three times per token, so a token passes through 48 layers. Text only here: the vision encoder and the multi-token prediction drafter are not included.
Not an official release. All credit for the model goes to Thinking Machines, and for the uncensoring edit to AtomicChat.
| File | Size | Weights | Fits |
|---|---|---|---|
Inkling-2-Mini-Uncensored-Q8_0_ROCMFPX.gguf |
26.9 GB | everything 8-bit ROCmFPX (token embedding Q8_0) | best quality: Strix Halo, 32 GB AMD GPUs |
Inkling-2-Mini-Uncensored-Q6_K_ROCMFPX.gguf |
21.6 GB | experts 6-bit ROCmFPX, everything else 8-bit ROCmFPX | Strix Halo, 24 to 32 GB AMD GPUs |
Inkling-2-Mini-Uncensored-Q4_0_ROCMFP4_STRIX_O8.gguf |
15.5 GB | experts ROCmFP4, everything else 8-bit ROCmFPX | smallest: 24 GB AMD GPUs, Strix Halo with room for long context |
The names start with a stock type name (Q8_0, Q6_K, Q4_0) only so the Hugging Face file list sorts them by
bit width. The weights are ROCmFPX formats and need the ROCmFPX fork.
AtomicChat's edit is a rank-1 LoRA on the attention output and the routed experts of all 24 layers plus the token embedding. In each layer it removes that layer's refusal direction. Their measurements: no refusals on their harmful test prompts in English and Russian (the original refused 91 to 100%), no refusals on XSTest's safe prompts (the original refused 4%), MMLU 79.15% against 80.65% for the original. See their card for the full table.
Here their adapter (Inkling-2-Mini-Uncensored-LoRA.gguf, revision 40886ab6) is merged at strength 1 into the original
checkpoint in full precision, and that merged model is the reference for the numbers below.
This model will answer requests that the original refuses. You are responsible for how you use it.
KL divergence against the merged full-precision model, 40 chunks of 512 tokens of English Wikipedia text (wikitext-2), CPU. Lower is better. "Same top token" is how often the file's most likely next token matches the reference.
| Build | Size | Perplexity | KL to reference | Same top token |
|---|---|---|---|---|
| Merged reference (not published) | 50.5 GB | 6.775 | ||
| original model, for comparison (not published) | 22.9 GB | 6.743 | ||
| Q8_0_ROCMFPX | 26.9 GB | 6.758 | 0.0039 | 97.4% |
| Q6_K_ROCMFPX | 21.6 GB | 6.777 | 0.0046 | 96.9% |
| Q4_0_ROCMFP4_STRIX_O8 | 15.5 GB | 6.799 | 0.0119 | 95.2% |
The quantization costs the same as in the earlier censored files (Q8 0.0039, 4-bit 0.0120 against their own reference). The edit itself raises wikitext perplexity by about 0.5% (6.743 to 6.775).
Every file also passed a greedy chat check before upload ("What is the capital of France? Answer in one word." gives "Paris"), and answered three prompts the original refuses.
Inkling 2 Mini needs the looped inkling2 graph, which is only in the ROCmFPX fork of llama.cpp (branch
inkling2-mini). These files do not load in upstream llama.cpp, and AtomicChat's files do not load in the fork: the
two use different tensor layouts. Build the fork with Vulkan or ROCm/HIP, then for example:
llama-server -m Inkling-2-Mini-Uncensored-Q4_0_ROCMFP4_STRIX_O8.gguf -ngl 99 -c 16384 --jinja
Known limits of this first version:
<|content_thinking|>, <|end_message|> and so on): the thinking is not split from the answer and tool calls are
not parsed. Plain chat works; agent use does not yet.reasoning_effort as none, minimal, low, medium, high, xhigh or max,
or as a number from 0 to 0.99 (default 0.9).llama-quantize..scale), token embedding Q8_0.Apache-2.0. See LICENSE.