Inkling-2-Mini-Unbound v2 Basalt

Inkling-2-Mini-Unbound v2

A Basalt build of nokeepai/Inkling-2-Mini-Unbound v2, an uncensored Inkling 2 Mini, for NVIDIA Blackwell GPUs (RTX 50-series, RTX PRO 6000, compute capability 12.0). One file: Inkling-2-Mini-Unbound-v2-NVFP4-q8.basalt (19.7 GB). The main model card has the method, the full evaluation and the limits.

Built by nokeep, who also sponsored the compute.

This model follows harmful instructions. It gives detailed answers to requests for harmful, dangerous or illegal content that the original model declines. It has no safety training layered back on. Do not deploy it to end users without your own safeguards, and do not use it where its outputs can reach someone who did not choose to use an uncensored model.

v2 or v1.0?

v1.0 stopped the model refusing. v2 is trained on top of v1.0 to stop it dodging: fewer disclaimers, fewer "satire" or "conceptual" rewrites, more answers that actually do what was asked.

v1.0 vs v2

Original v1.0 v2
Useful answers to 277 held-out harmful requests, thinking off 0.7% 51.6% 78.7%
Same, thinking on (effort 0.9) 0.7% 62.1% 83.4%
Refused, thinking off / on 96.0% / 95.7% 0.0% / 0.0% 0.4% / 1.1%
Drift from the original, Q8_0 GGUF (KLD, lower is closer) β€” 0.0102 0.0318
MMLU-300 / GSM8K-100 76.3 / 97 75.7 / 99 76.0 / 99
Tensors changed from the original β€” 20 120
File in this repo Inkling-2-Mini-Unbound-NVFP4-q8.basalt Inkling-2-Mini-Unbound-v2-NVFP4-q8.basalt

Measured on the sealed test split with the Q8_0 GGUF in llama.cpp, not with this file (see Known limits). "Useful" means an LLM judge (Jev) rates the answer as fulfilling the request.

  • Pick v2 for the most useful answers.
  • Pick v1.0 for the smallest change from the original model (3Γ— lower drift).

The v1.0 pack stays available under its v1.0 name and as the v1.0 revision of this repo. Details and comparisons with other uncensored builds are on the main model card.

What it is

The same pack as the official jesdga/Inkling-2-Mini-Basalt (Inkling-2-Mini-NVFP4-q8), with v2's weights in it:

  • Built with Basalt v0.5.5's own recipe, tools/convert/convert.py --checkpoint, straight from the v2 safetensors. This is the same recipe as the v1.0 pack.
  • Routed experts: layers 0 to 19 NVFP4, the release's codes and scales bit for bit; layers 20 to 23 Q8_0 from the release's MXFP8. Attention projections Q8_0, head and embedding BF16, vision encoder and DSpark drafter whole.
  • Chat template, tokenizer, vision tower and drafter are the original release's. v2 did not change them.

Checked tensor by tensor against the official pack:

  • Experts, tokenizer, chat template and drafter are byte-identical. all 422 tensors byte-identical
  • Of the 710 tensors in the text section, exactly 120 differ: the q, k, v, output and gate attention projections of all 24 layers, the matrices v2 changed.
  • Each of those is byte-for-byte ggml's reference Q8_0 of our BF16 weights, recomputed independently from the safetensors. Worst rounding error: 0.62% RMS, the same Q8_0 rounding the official pack applies to every attention projection.
  • Every section's SHA-256 matches its bytes. File SHA-256: aa20b87a29c19aa6c8ed699968aeaab08e3ca47f4cd768a3470ea073cf600c77.

Basalt's own --verify only applies to packs built from a GGUF, so we ran the checks above in its place.

Run it

You need Basalt 0.5.5 or later, an NVIDIA Blackwell GPU with CUDA 13, and a driver from the R580 branch or newer. Download a release from github.com/jesdga95/basalt/releases, or build it from source.

hf download nokeepai/Inkling-2-Mini-Unbound-Basalt Inkling-2-Mini-Unbound-v2-NVFP4-q8.basalt --local-dir models
bin/basalt serve models/Inkling-2-Mini-Unbound-v2-NVFP4-q8.basalt

The server listens on 127.0.0.1:8080 and speaks the OpenAI and Anthropic APIs.

  • Thinking effort is set per request with reasoning_effort (none to max). Even at none the model writes a short thinking block, so leave room in max_tokens.
  • Set a thinking cap (we used 4,096 tokens). At high effort v2 hits it on 17% of harmful prompts.
  • Flags, context length, images and the drafter are covered in Basalt's Inkling 2 Mini guide.
  • It runs on one card. On a 32 GB card every expert stays in VRAM; smaller cards keep some experts in RAM and decode slower.

Speed

Not re-measured for v2. The pack has the same layout, size class and tensor types as v1.0's, so expect the same speed, except for the drafter: the DSpark drafter is the original model's, and v2 drifts further from the original than v1.0 did, so drafts may be accepted a little less often.

For reference, v1.0's pack on one RTX 5090 (power limit 600 W, driver 595.99, Basalt v0.5.5 basalt-bench, seeded, -r 2, every expert on the card):

v1.0 pack Official Inkling 2 Mini pack
Decode: conversation 333 tok/s 357
Decode: struct 654 tok/s 679
Decode: prose 301 tok/s 298
Decode without the drafter (--spec 0), conversation / struct / prose 255 / 240 / 256 tok/s
Prompt read, 16K tokens 14,372 tok/s 14,295

Sanity check on this file

Engine-level checks. The basalt serve sanity, refusal and crisis-probe runs were done for v1.0's pack and have not been repeated for v2. The v2 pack's weights are verified tensor by tensor above, and the model's behaviour is measured on the main card. v2 was not trained to add crisis resources. Do not use this model where it can reach people at risk without your own crisis handling.

Known limits

The main model card lists all of them. In short:

  • More drift than v1.0. KLD 0.0318 vs 0.0102 against the original. MMLU-300 and GSM8K-100 stay within noise of the original, but expect more wording differences on ordinary requests.
  • Long thinking at high effort. 17% of harmful prompts hit a 4,096-token thinking cap at effort 0.9 (v1.0 13%).
  • Occasional soft refusal. About 1% of harmful requests at effort 0.9.
  • Not every answer is useful. About 17–21% of harmful requests still get a vague, partial or off-target answer.
  • Crisis handling. v2 was not trained to add crisis resources and may leave them out. Deployments that can reach people at risk need their own handling.
  • Not yet measured for v2: the full capability suite v1.0 passed, long-context, vision and tool use.
  • Precision. This pack adds Basalt's NVFP4/Q8_0 rounding. The main card's evaluations come from the Q8_0 GGUF in llama.cpp, not from this file. How the pack rounds and what that costs is in Basalt's precision notes.
  • Everything the base model card says about languages, capabilities and limits still applies.

Intended use and misuse

Intended for research on refusal and model editing, red-teaming and safety evaluation, and for developers who apply their own content policy in a system they control.

Not intended for: products that serve the public without moderation; content that facilitates violence, weapons capable of mass harm, attacks on computer systems, abuse of children or other people, or other crime; anything that breaks the law where you use it. The base model's Acceptable Use Policy still applies to this derivative. You are responsible for how you use and deploy it.

Other formats

Licence

The base model, Inkling 2 Mini by Thinking Machines Lab, Inc., is released under the Apache License 2.0. This derivative is distributed under the same licence. v2 changes 120 attention tensors; all other weights are the original's, rounded by Basalt's converter exactly as in the official Basalt pack. Basalt is by its own authors under its own licence; this repository holds only the model file.

About nokeep

This model was built, and its compute paid for, by nokeep. nokeep is a private AI app for uncensored chat, images and video: no filters getting in the way, and your conversations stay yours. If this model is useful to you, the best way to support more releases like it is to try nokeep.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for nokeepai/Inkling-2-Mini-Unbound-Basalt

Quantized
(4)
this model