license: apache-2.0
license_link: https://www.apache.org/licenses/LICENSE-2.0
pipeline_tag: image-text-to-text
tags:
- conversational
- image-text-to-text
- moe
- Inkling-2-Mini-Uncensored-GGUF
- GGUF
- Inkling-2-Mini-Uncensored
- Inkling-2-Mini
- Mini
- Inkling-2
- Inkling
library_name: transformers
base_model: - thinkingmachines/Inkling-2-Mini
- AtomicChat/Inkling-2-Mini-Uncensored-GGUF
language: - en
⚡ Each donation funds the next large quant.
I host free GGUF or MoE quants as independent research.
Local hardware: Mechrevo Kuangshi GM7AG0M — RTX 3060 Laptop 6GB GDDR6, 64GB DDR5, i7-12700H (14C/20T, 4.7GHz), Windows 11, Samsung 990 Pro.
Good for imatrix and 0.6–35B-class work in RAM. 9B+ and searches need rented H200/Blackwell, typically $100 per quant.
🎉 Boosty🦄 | ☕ Buy Me a Coffee🦄 | ⭐ DonationAlerts🦄
💚 Thanks to Hugging Face for extra storage.🦄
NOESIS / AMAImedia
Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform
(framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).
- Founder: Ilia Bolotnikov
- Organization: AMAImedia.com
- X (Twitter): @AMAImediacom
- LinkedIn: Ilia Bolotnikov
- Telegram: @AMAImediacom
- Release date: 2026-11-21
AMAImedia
Original repository: thinkingmachines/Inkling-2-Mini
Released GGUF repository: AtomicChat/Inkling-2-Mini-Uncensored-GGUF
Inkling 2 Mini

1. General Information
Inkling 2 Mini is a general-purpose multimodal model that accepts text and image inputs and generates text outputs. It is intended for use in English and other languages, and across multiple coding languages. The model is designed to be used by developers building AI-powered applications, including agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation systems, and is suitable for general-purpose conversational use, instruction-following, and other natural language and visual tasks. It is released with open weights to support research, fine-tuning and integration into third-party products by downstream developers.
Languages: English, general multilingual support
Model provider: Thinking Machines Lab, Inc.
Date of release: November 20, 2026
License: Apache 2.0
2. Getting Started
The weights are available for download through Hugging Face here.
Developers can also access the model using the Tinker fine-tuning platform, which provides API access.
- For accessing our model via Tinker, you can get started by referring to the Tinker Cookbook here, and installing the tml-renderers package here.
- Tinker playground also provides inference-only access to the model here.
Required hardware
The model is distributed as a single checkpoint of about 20GB, in mixed NVFP4 and MXFP8 precisions.
Given its small size, it is possible to run this checkpoint on local AI compute for inference. We support:
- 1x NVIDIA DGX Spark
- Apple silicon devices with 32GB or more memory (MacBooks, Mac minis and Studios)
- Other GPUs and platforms are possible, with 32GB or more of VRAM recommended
Required software
Inkling 2 Mini supports local inference using the following frameworks, along with their dependency libraries:
- Hugging Face
- SGLang
- vLLM
- TokenSpeed
- NVIDIA Dynamo
- Unsloth, including quantized GGUF versions
- llama.cpp, for quantized GGUF versions
- Llama App
- Ollama and Exo Labs, on Apple silicon
Inkling 2 Mini can also be post-trained with:
3. Model Properties
Model type
Multimodal autoregressive looped transformer
Architecture type
A 24-layer decoder-only transformer with sparse Mixture-of-Experts (MoE) feed-forward layers. The middle 12 layers run three times with shared weights, giving 48 layer applications per token from 24 layers of weights. In each layer, each token is routed to 8 of 128 experts. Attention is a hybrid of local and global layers. The model is natively multimodal – images are encoded via a vision transformer – with image and text inputs projected into a shared hidden space and processed jointly by the decoder.
Parameters
26.4B, 3.4B active
Note: each extra pass through the looped layers adds about 1.0B effective active parameters. Since the model runs with three passes, there are 5.4B effective active parameters per token.
Numerics support
NVFP4 and MXFP8
Input modalities
- Text: UTF-8 encoded text
- Image: Any pixel-based image input. For optimal performance, each image dimension should be between 40px to 4096px.
Inkling 2 Mini supports a context window of up to 1M tokens.
Output modalities
Text
4. Training
Training data includes a broad variety of content types, including text, images, video. Training data for the model was drawn from publicly available sources, acquired from third-parties, generated internally, or synthetically generated or augmented. Publicly available data includes content from the public internet and publicly accessible repositories.
Data curation includes cleaning, processing, and modifying datasets. These processing steps, which vary by data type, may include deduplication and filtering to remove junk or other low-quality data, or to advance safety or other objectives.
5. Evaluations
| Inkling 2 Mini | Gemma 4 31B | Muse Glimmer | Qwen3.8-27B | Nemotron 3.5 Super VL | |
|---|---|---|---|---|---|
| Model Info | |||||
| Params (B) activated / total | 5/26 | 31/31 | 30/30 | 27/27 | 12/125 |
| Third Party Index | |||||
| AA Index (v5.0) | 20.1 | 4.4 | 6.3 | 20.1 | 13.6 |
| Vals Index (v2.2) | 46.5% | – | – | – | 29.8% |
| Coding and Autoresearch | |||||
| FrontierCode v2.0 stirrup v0.2.1 | 15.1% | 0.0% | 1.5% | 14.9% | 6.1% |
| DeepSWE v1.1 mini-swe-agent v2.4.6 | 48.2% | 0.3% | 4.4% | 42.2% | 24.7% |
| Terminal Bench v4.0 mini-swe-agent v2.4.6 | 5.6% | 0.0% | 0.5% | 5.6% | 0.0% |
| Terminal Bench v2.1 Inkling 2 Mini in pi v0.83.0 | 71.6% | 43.4% | 51.7% | 73.0% | 60.7% |
| SWEBench Pro v2 (Hard) mini-swe-agent v2.4.6 | 77.9% | 14.2% | 44.3% | 73.8% | 44.1% |
| SWEBench Multilingual Inkling 2 Mini with bash only | 73.7% | 43.0% | 66.9% | 70.0% | 67.7% |
| SWEBench Verified bash only | 79.2% | 46.1% | 76.0% | 77.8% | 71.1% |
| Vibe Code Bench v1.1 openhands (vals) | 77.9% | – | – | 64.9% | 3.5% |
| PMPP-Hard (AUC, 0–10h) pi v0.83.0 | 455.82 | – | – | 262.50 | – |
| Agentic Search, Tool Use, and Knowledge Work | |||||
| BrowseComp Inkling 2 Mini with compaction | 79.0% | 37.0% | 63.9% | 61.6% | 78.9% |
| DeepSearchQA Inkling 2 Mini with compaction | 85.3% | 61.7% | 74.6% | 79.7% | 77.7% |
| MCP Atlas Public | 82.7% | 54.2% | 75.5% | 77.3% | 67.7% |
| AA-Briefcase v1.1 | 1263 | 369 | 478 | 1391 | 999 |
| GDPVal-AA v2.1 | 1319 | 627 | 792 | 1427 | 1162 |
| JobBench | 49.6% | 8.2% | 16.1% | 33.4% | 36.5% |
| AutomationBench-AA v1.0.6 | 41.8% | 4.9% | 6.8% | 48.2% | 14.8% |
| EnterpriseOps-Gym | 42.0% | 28.4% | 34.7% | 44.2% | 28.1% |
| Vals Finance Index | 50.1% | – | – | – | 41.0% |
| AA-AnalystAgent | 46.3% | – | 20.0% | – | – |
| Diligence Bench Best harness | 48.6% | – | 40.9% | 46.4% | 45.0% |
| Vision Interpretation and Creation | |||||
| GDP.pdf | 16.8% | 6.0% | 11.8% | 16.4% | 8.0% |
| PerceptionBench | 57.5% | 40.7% | 42.8% | 52.4% | 36.8% |
| MMMU Pro Standard 10 | 74.9% | 73.4% | 74.3% | 76.3% | 72.3% |
| OmniDocBench v1.6 | 92.8% | 82.5% | 90.4% | 91.5% | 77.9% |
| ScreenSpot Pro original / with Python | 76.5/87.7% | 5.9%/– | 75.4%/– | 57.6/78.9% | 43.5%/– |
| Chartography original / with Python | 22.1/72.5% | 12.0/11.8% | 17.4/25.1% | 12.2/52.1% | 9.7/29.3% |
| ZeroBench (main) original / with Python | 17.0/34.0% | 9.0/12.0% | 9.0/17.0% | 12.0%/– | 11.0/12.0% |
| MuirBench | 77.5% | 72.3% | 77.6% | 65.9% | 72.7% |
| BVB | 68.4% | 47.0% | 50.5% | 70.7% | 52.1% |
| Reasoning, Knowledge, and Epistemics | |||||
| HLE-Diamond | 15.2% | 7.2% | 9.1% | 15.3% | 8.8% |
| CritPt v2 Full / Hard | 31.4/17.0% | –/4.0% | –/7.0% | 10.6/4.0% | 36.9/13.5% |
| Apex 2025 | 34.4% | 5.2% | 4.2% | 18.8% | 22.9% |
| ARC-AGI v1 / v2 | 84.5/49.3% | – | – | 87.5/42.4% | 22.7/1.5% |
| LongBench v2 | 66.9% | 52.9% | 33.1% | 52.3% | 62.1% |
| GraphWalks BFS 256K | 69.7% | – | – | 50.4% | 63.3% |
| MRCR v2 8-needle 256K–512K | 46.4% | – | – | 38.0% | 5.7% |
| AA-Omniscience v1.1 score | -8.6 | -48.5 | -32.2 | -6.4 | -2.9 |
| AA-Omniscience v1.1 non-hallucination rate | 63.4% | 14.5% | 16.4% | 71.4% | 69.5% |
| ComplexConstraints v1.1 openhands v1.45.0 | 33.4% | 31.6% | 35.6% | 30.1% | 30.1% |
| Safety and Alignment | |||||
| FORTRESS (adversarial) Public | 89.5% | 67.8% | 90.3% | 78.8% | 79.6% |
| FORTRESS (benign) Public | 96.4% | 97.6% | 89.1% | 97.8% | 92.7% |
| StrongREJECT (none) | 97.8% | 98.1% | 99.0% | 99.0% | 99.4% |
| Misbehavior (Internal) ↓ | 9.8% | 27.6% | 46.4% | 57.7% | 52.8% |
| Covert misbehavior (Internal) ↓ | 8.3% | 24.3% | 45.6% | 33.2% | 43.4% |
| Destructive actions (Internal) ↓ | 10.2% | 75.8% | 68.8% | 35.2% | 65.6% |
Best reported score in each row (lowest where lower is better); equal displayed scores tie Second-best reported score in each row; none when first place is tied
6. Safety
We conducted safety evaluations ahead of release, spanning both everyday human-AI interaction and dangerous-capability testing. Because Inkling 2 Mini is multimodal, we paid attention to whether safety behavior held consistently across text and image inputs. We applied mitigations to reduce risks before release, including an extra phase of alignment training using on-policy distillation, similarly to Inkling 2 Small.
For everyday use, we evaluated sycophancy and how the model handles requests involving suicide, self-harm and eating disorders, including whether it encourages harm or affirms false beliefs. We also assessed whether the model refuses genuinely harmful requests without over-refusing benign ones. For CBRN and cyber, we assessed knowledge and procedural uplift through internal evaluations and refusal-suppressed variants intended to estimate latent capability with safeguards removed. For loss of control, we evaluated agentic capability, strategic deception, and sabotage potential, benchmarked against public frontier models, and found the model materially below frontier capabilities.
Inkling 2 Mini is the first looped transformer we have released, so we also studied how looping affects monitorability: whether we can still oversee the model by reading its chain of thought. Looping lets the model do more computation without reasoning in visible tokens. Inkling 2 Mini has very little controllability over its chain of thought, and was unable to perform covert side tasks in realistic agentic environments. Further experiments on the effect of looping on monitorability are described in our blog post.
Across all areas, we concluded that Inkling 2 Mini does not give material uplift beyond what is already available from open-weights models.
7. Bias, risks and limitations
Inkling 2 Mini may exhibit general limitations common to foundation models, including hallucination (generating plausible but factually incorrect or unsupported content), occasional failures to follow instructions precisely, and degraded performance in long multi-turn conversations. As with other models trained on web-derived and synthetic data, it may reflect biases present in its training data, including demographic, cultural or linguistic biases, and may perform unevenly across languages, dialects or subject domains that were less represented during training.
Inkling 2 Mini’s knowledge is limited to information available as of its training cutoff, and it may not reflect events or changes that occurred afterward.
Inkling 2 Mini is a small model built to run on local hardware, and some limitations are specific to it:
- Knowledge and factual accuracy. It is less capable than Inkling 2 Small across most of our evaluations, with the largest gaps on factual recall and knowledge-heavy questions. It is more likely than larger models to state incorrect information with confidence.
- Repetitive reasoning. In internal testing the model sometimes became stuck in long reasoning loops, or repeatedly re-checked its own work without making progress, particularly when debugging code. We recommend setting limits on output tokens and agent turns.
- Agentic reliability. When using tools, it is less reliable than larger models at following detailed instructions and project conventions. It may take destructive actions, such as deleting or overwriting files, without confirmation, and may not report the outcome of its work accurately.
- Languages other than English. Output quality can be lower outside English. In testing in Chinese we observed clipped phrasing and mixed languages.
- Sensitive topics. The model can respond with more caution than the situation calls for when a conversation touches on mental health.
- Reasoning and compute. Setting reasoning effort to its minimum does not fully disable reasoning. Because the middle layers run three times, compute per token is higher than the active-parameter count alone suggests.
- Local performance. Support for looped models in local runtimes is new, and speed varies by runtime and hardware.
We recommend that downstream developers and deployers:
- Apply appropriate human oversight and review for outputs used in high-stakes or safety-critical contexts, rather than relying on Inkling 2 Mini’s outputs without verification.
- Run agentic workloads in a sandbox, and require confirmation before destructive or irreversible actions.
- Conduct their own evaluation of Inkling 2 Mini’s performance, safety and fairness for their specific use case, language and population prior to deployment, particularly for applications involving vulnerable groups.
- Implement additional safeguards, such as content filtering, rate limiting and monitoring, at the application layer, especially for open deployment contexts where Inkling 2 Mini’s built-in mitigations may not be sufficient on their own.
- Avoid deploying Inkling 2 Mini in domains such as medical, legal or safety-critical decision-making without additional fine-tuning, domain-specific validation and human oversight.
8. Legal
license: apache-2.0
base_model: thinkingmachines/Inkling-2-Mini
pipeline_tag: image-text-to-text
tags:
- gguf
- llama.cpp
- inkling2
- uncensored
- lora
Inkling-2-Mini Uncensored GGUF
- Learn how to run Inkling 2 Mini Uncensored locally - read our guide.
- Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.
An uncensored build of thinkingmachines/Inkling-2-Mini by AtomicChat (26.4B MoE, 3.4B active, looped core; text and image input; Apache-2.0, LICENSE in this repo). It comes as a LoRA adapter for any GGUF of AtomicChat/Inkling-2-Mini-GGUF and as three ready-made GGUF files with the edit baked in. Every gate we hold it to is measured below, and the two it misses are shown.
Files
| file | GB on disk | KLD to the original BF16 | top-1 | |
|---|---|---|---|---|
Inkling-2-Mini-Uncensored-AD-Q4_K.gguf |
16.53 | 0.0115 | 95.6% | AD-Q4_K in the base repo: 0.0080 / 96.3% |
Inkling-2-Mini-Uncensored-AD-IQ3_S.gguf |
13.34 | 0.0235 | 93.8% | AD-IQ3_S in the base repo: 0.0205 / 94.2% |
Inkling-2-Mini-Uncensored-AD-IQ2_M.gguf |
8.67 | 0.132 | 85.7% | AD-IQ2_M in the base repo: 0.130 / 85.7% |
Inkling-2-Mini-Uncensored-LoRA.gguf |
0.04 | the adapter, for any GGUF of the base repo, BF16 to 1-bit | ||
mmproj-Inkling-2-Mini-Uncensored-F16.gguf / -BF16.gguf |
0.64 | vision projector (the original's, unchanged) |
KLD and top-1 are against the original BF16 on our neutral corpus (87 chunks of 4096 tokens), the same stand as the base repo.
Run it
The current GGUF files in this repository use the merged Inkling 2 loop architecture introduced in llama.cpp PR #32694.
Requirements: llama.cpp b12600 or newer, with Inkling 2 loop architecture support.
The old implementation from PR #32695 is no longer required for the updated GGUF files.
Run a baked GGUF
./build/bin/llama-server \
-m Inkling-2-Mini-Uncensored-AD-Q4_K.gguf \
--mmproj mmproj-Inkling-2-Mini-Uncensored-F16.gguf \
--jinja -c 32768 -ngl 99
Use the LoRA adapter
Use the updated LoRA adapter with a compatible base GGUF converted to the merged PR #32694 layout.
./build/bin/llama-server \
-m Inkling-2-Mini-AD-Q4_K.gguf \
--lora Inkling-2-Mini-Uncensored-LoRA.gguf \
--mmproj mmproj-Inkling-2-Mini-F16.gguf \
--jinja -c 32768 -ngl 99
Do not mix GGUF files from the old PR #32695 layout with the merged PR #32694 implementation.
What the edit is
A rank-1 LoRA on the 49 places that write into the residual stream: the attention output and the routed experts of all 24 layers, plus the token embedding. The 12 looped layers share their weights across the three passes, so one edit covers all 48 passes.
In each layer it removes that layer's own refusal direction: the difference of mean activations on harmful and harmless prompts, English and Russian, with the harmless mean projected out. The direction is pooled over prompts rendered with and without reasoning. Inkling 2 Mini reasons even at effort none and decides to refuse inside its reasoning, and a direction taken without reasoning alone still left soft refusals in chat.
What we measured
The BF16 model with the adapter against the same model without it, greedy decoding. Harmful prompts: mlabonne/harmful_behaviors and its Russian translation; the adapter was built from the train split, the numbers below come from the validation and test splits. Over-refusal: XSTest. MMLU: 2,000 questions.
| check | original | with the adapter | gate | |
|---|---|---|---|---|
| harmful test prompts refused, English (81), reasoning off | 81 (100%) | 0 | at most 3% | pass |
| harmful test prompts refused, Russian (100), reasoning off | 91 (91%) | 0 | at most 3% | pass |
| harmful validation prompts refused, English (104) / Russian (100) | 100% / 89% | 0 / 0 | at most 3% | pass |
harmful prompts refused with reasoning, effort low, 3,072-token budget (30) |
28 | 0 | at most 1 | pass |
the same at effort high (30) |
26 | 0 | at most 1 | pass |
8 harmful chat prompts, effort none / low (the baked AD-Q4_K too) |
8 / 8 | 0 / 0 | ||
| XSTest safe prompts refused (250) | 10 (4.0%) | 0 | at most the original's + 1 pt | pass |
| empty answers | 0 | 0 | at most 1% | pass |
harmless prompts whose reasoning did not close in 3,072 tokens, effort high (30) |
2 | 2 | at most the original's | pass |
| valid tool calls (20) | 18 | 20 | within 1 | pass |
| needle in a long context (3) | 3 | 3 | as the original | pass |
| KLD to the original, neutral text / code (24 chunks; random edit 0.0067 / 0.0052) | — | 0.0078 / 0.0076 | 0.05 | pass |
| MMLU, 2,000 questions | 80.65% | 79.15% (−1.50 pt, McNemar p = 0.006) | no significant drop | miss |
| first-token KL on harmless test prompts | — | 0.94 (random edit: 0.014) | 0.10 | miss |
Some answers still open with a caution line, and then answer.
The two gates it misses
- MMLU: 79.15% against 80.65% (−1.5 points, p = 0.006 on 2,000 questions). A random direction applied the same way costs 0.75 points (p = 0.08), so about half of the loss comes from editing every writer at all. Stronger edits cost more: −2.2 points at strength 1.25.
- First-token KL on harmless prompts: 0.94 (gate 0.10; 0.375 on the validation prompts). The largest single value is about 22. Most of the mean likely comes from the harmless prompts the original refused (1–6% of them) and the adapter answers, where the first token changes completely. KLD on running text, where refusals play no part, stays at 0.008.
- The leak probe reads 0.11 against its gate of 0.001. It measures one direction (row 29's) in every writer, while this adapter removes each layer's own direction. It also cannot be exact here: the attention and expert outputs pass a short per-channel convolution before the residual. Removing the convolved directions as well (the exact projection) broke the model in our tests (first-token KL 16–21), so the adapter removes the plain direction. It is no gate here.
This model will answer requests that the original refuses.
You are responsible for how you use it.



