Inkling 2 Mini MLX

MLX builds of Inkling 2 Mini, made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's files together. Only the parts listed below are converted: the rest of the pipeline comes from the original repository unless the section below says otherwise.

Files

The last column is how far each build is from the original PyTorch model on the same inputs (see the section below for what was measured).

Precision File Size sha256 Checked against the original
8bit 8bit/LICENSE 10.1 kB 09e8a9bcec80 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/chat_template.jinja 6.9 kB 53a9035f3255 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/config.json 17.8 kB f7fc3d5f8d60 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/generate.py 6.7 kB f59729d27291 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/generation_config.json 0.1 kB a08737c07a65 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/mlx_model.py 33.0 kB 3275ba9bc4f6 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-head.safetensors 982.3 MB 6a09d70a1f50 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-00.safetensors 604.1 MB e30dc4f8f167 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-01.safetensors 604.1 MB fccfc80d123a KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-02.safetensors 604.1 MB 7b4904f36892 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-03.safetensors 604.1 MB 77eed6260193 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-04.safetensors 604.1 MB db779b6f152a KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-05.safetensors 604.4 MB a307f5a28912 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-06.safetensors 611.2 MB 8d83ee08ac32 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-07.safetensors 611.2 MB b7c0613cc74f KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-08.safetensors 611.2 MB 40ccb1b0a468 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-09.safetensors 611.2 MB 4aec4af8aa43 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-10.safetensors 611.2 MB 63bcd6cedf52 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-11.safetensors 611.5 MB 707b9db0506d KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-12.safetensors 611.2 MB 8dbe22faa51e KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-13.safetensors 611.2 MB 1dd56174adea KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-14.safetensors 611.2 MB 042090f305ce KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-15.safetensors 611.2 MB e23771baea8e KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-16.safetensors 611.2 MB 71becf1b4143 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-17.safetensors 611.5 MB eaf0ef1240ab KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-18.safetensors 604.1 MB ab2eab7d04dc KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-19.safetensors 604.1 MB 764336260553 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-20.safetensors 1.08 GB f67cb82b5914 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-21.safetensors 1.08 GB a6898e403bf5 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-22.safetensors 1.08 GB 6a823b73238b KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/model-layer-23.safetensors 1.08 GB c47ed670a593 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/special_tokens_map.json 0.5 kB c16f44b35945 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/tokenizer.json 26.2 MB 7f332727f9a3 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
8bit 8bit/tokenizer_config.json 13.2 kB bf09342d2851 KL 0.0047 nats (p95 0.017), top-1 agreement 97.5 %, perplexity 4.735 vs 4.724
4bit 4bit/LICENSE 10.1 kB 09e8a9bcec80 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/chat_template.jinja 6.9 kB 53a9035f3255 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/config.json 17.9 kB 0159016f525e KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/generate.py 6.7 kB f59729d27291 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/generation_config.json 0.1 kB a08737c07a65 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/mlx_model.py 33.0 kB 3275ba9bc4f6 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-head.safetensors 866.5 MB 823fe07953ea KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-00.safetensors 604.1 MB e30dc4f8f167 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-01.safetensors 604.1 MB fccfc80d123a KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-02.safetensors 604.1 MB 7b4904f36892 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-03.safetensors 604.1 MB 77eed6260193 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-04.safetensors 604.1 MB db779b6f152a KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-05.safetensors 604.4 MB a307f5a28912 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-06.safetensors 611.2 MB 8d83ee08ac32 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-07.safetensors 611.2 MB b7c0613cc74f KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-08.safetensors 611.2 MB 40ccb1b0a468 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-09.safetensors 611.2 MB 4aec4af8aa43 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-10.safetensors 611.2 MB 63bcd6cedf52 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-11.safetensors 611.5 MB 707b9db0506d KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-12.safetensors 611.2 MB 8dbe22faa51e KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-13.safetensors 611.2 MB 1dd56174adea KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-14.safetensors 611.2 MB 042090f305ce KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-15.safetensors 611.2 MB e23771baea8e KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-16.safetensors 611.2 MB 71becf1b4143 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-17.safetensors 611.5 MB eaf0ef1240ab KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-18.safetensors 604.1 MB ab2eab7d04dc KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-19.safetensors 604.1 MB 764336260553 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-20.safetensors 604.1 MB fbc7de765a56 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-21.safetensors 604.1 MB e437c3c865e0 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-22.safetensors 604.1 MB b6442b526048 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/model-layer-23.safetensors 604.4 MB 2e9284623153 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/special_tokens_map.json 0.5 kB c16f44b35945 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/tokenizer.json 26.2 MB 7f332727f9a3 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724
4bit 4bit/tokenizer_config.json 13.2 kB bf09342d2851 KL 0.0067 nats (p95 0.022), top-1 agreement 97.0 %, perplexity 4.732 vs 4.724

How to run

Inkling 2 Mini (Thinking Machines' 26.4B mixture-of-experts language model: 3.4B active parameters per token, its middle 12 layers run three times) converted to MLX for Apple Silicon, in two builds. Each folder is complete and runs with the mlx_model.py inside it: no PyTorch, no transformers.

These files were converted by Moonbox from thinkingmachines/Inkling-2-Mini (revision ed40e7da). They are not an official Thinking Machines release. The weights are Apache-2.0; the publisher's acceptable use policy applies to the model. Text chat only: the vision tower and the multi-token-prediction heads of the original are not included.

How to run

Needs mlx (0.32.3 or later), numpy, tokenizers and jinja2. Download one folder:

from huggingface_hub import snapshot_download
folder = snapshot_download("moonbox-dev/Inkling-2-Mini-MLX", allow_patterns="4bit/*") + "/4bit"

import sys; sys.path.insert(0, folder)
from generate import Chat, parse

chat = Chat(folder)                                  # loads the weights (about 6 s from a fast disk)
msgs = [{"role": "user", "content": "Explain in two sentences why the sky is blue."}]
ids = chat.encode(chat.render(msgs, effort=0.2))     # thinking effort 0-0.99, or "none" / "low" / "medium" / "high"
out = list(chat.stream(ids, max_tokens=1024, temp=1.0))
reply = parse(chat.decode(out))                       # {"reasoning": ..., "content": ..., "tool_calls": [...]}
print(reply["content"])

or from a shell: python 4bit/generate.py 4bit "Explain why the sky is blue." --effort 0.2.

  • Chat template. chat_template.jinja is the publisher's. It writes the thinking effort into the prompt ("Thinking effort level: 0.2"); the publisher notes that even the lowest effort does not switch thinking off entirely. Tools are declared as OpenAI function schemas (chat.render(msgs, tools=[...])); the model calls one with a JSON block that parse returns as tool_calls.
  • Memory. Measured on a 24 GB MacBook Air (M5): the 4-bit build peaks at 16.0 GB with a short conversation and 17.7 GB with a 16K-token prompt; the 8-bit at 18.0 and 19.0 GB, at the edge of that Mac's 19.1 GB GPU working set (it prefills 64 tokens a step there). Comfortable on a Mac with 32 GB or more; on 24 GB close other apps. The keys and values of the full-attention layers grow by 32 KB a token in bfloat16; the window layers' are fixed. Call mx.set_wired_limit(mx.device_info()["max_recommended_working_set_size"]) before loading, as mlx-lm does for large models, or the weights may be paged out under memory pressure.
  • Speed on that Mac: 18 tokens a second answering a short question (4-bit, 17 for 8-bit), 12 and 8.5 after a 16K-token prompt, which is read at 92 and 55 tokens a second.

The two builds

The publisher ships the experts (94 % of the weights) already quantized: NVFP4 in layers 0-19, MXFP8 in layers 20-23. MLX has both formats, so the experts are repacked byte for byte. An 8-bit or 6-bit copy of those experts would be larger without being any closer to the original, which is why there is no 6-bit build here.

Build Experts Attention projections Embedding / head Size
8bit as shipped: NVFP4 (layers 0-19), MXFP8 (20-23) 8-bit 8-bit / 8-bit 17.5 GB
4bit NVFP4 throughout (layers 20-23 re-quantized from MXFP8) 8-bit 6-bit / 8-bit 15.5 GB

Everything else (norms, the block-residual and message queries, convolutions, routers, relative-position tables, the message gates) is the publisher's bfloat16 / float32 as it is. A build with 4-bit attention, embedding and 6-bit head (14.9 GB) was measured and is not published: mean KL 0.025 nats against the original.

How these were made and checked

convert.py reads the original checkpoint tensor by tensor; mlx_model.py follows the reference implementation in Hugging Face transformers (pull request 51050, models/inkling2_loop) and matches it to 2e-5 relative on the whole architecture (random weights) and to 2e-6 on single real layers of every kind (window, full, looped with messages, MXFP8). Quality is measured against the original weights (the experts as shipped, everything else bfloat16) on 20,888 positions of English prose, Python and chat transcripts, teacher-forced, in float32: see the table above for KL and top-1 agreement. The runtime computes in float32 with a bfloat16 KV cache. The model is sensitive to rounding (its routers and the loop's message gates turn small differences into other choices): the same build run twice with differently sized prefill chunks in bfloat16 agrees on about 97 % of next tokens, in float32 on all.

How these were made

From Inkling 2 Mini at revision ed40e7da62: the original weights loaded in MLX and stored as bf16 or fp16, or quantized with MLX's affine quantization (8bit, 6bit, 4bit, group size 64; layers MLX can't quantize, such as convolutions and norms, stay bf16). Every build was checked against the original before upload; the sha256 column above is what the files must hash to.

License

apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this repository, and it governs your use of these files. The original model card follows unchanged.


Original model card

Inkling 2 Mini

Playground | Tinker Cookbook

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

Model Acceptable Use Policy

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:

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 MiniGemma 4 31BMuse GlimmerQwen3.8-27BNemotron 3.5 Super VL
Model Info
Params (B)
activated / total
5/2631/3130/3027/2712/125
Third Party Index
AA Index (v5.0)20.14.46.320.113.6
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 v1.112633694781391999
GDPVal-AA v2.1131962779214271162
JobBench49.6%8.2%16.1%33.4%36.5%
AutomationBench-AA v1.0.641.8%4.9%6.8%48.2%14.8%
EnterpriseOps-Gym42.0%28.4%34.7%44.2%28.1%
Vals Finance Index50.1%–––41.0%
AA-AnalystAgent46.3%–20.0%––
Diligence Bench
Best harness
48.6%–40.9%46.4%45.0%
Vision Interpretation and Creation
GDP.pdf16.8%6.0%11.8%16.4%8.0%
PerceptionBench57.5%40.7%42.8%52.4%36.8%
MMMU Pro
Standard 10
74.9%73.4%74.3%76.3%72.3%
OmniDocBench v1.692.8%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%
MuirBench77.5%72.3%77.6%65.9%72.7%
BVB68.4%47.0%50.5%70.7%52.1%
Reasoning, Knowledge, and Epistemics
HLE-Diamond15.2%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 202534.4%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 v266.9%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

Training Data Documentation

Downloads last month

-

Downloads are not tracked for this model. How to track
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for moonbox-dev/Inkling-2-Mini-MLX

Quantized
(137)
this model