File size: 5,134 Bytes
82141c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
---
license: apache-2.0
base_model: thinkingmachines/Inkling-2-Mini
pipeline_tag: text-generation
library_name: mlx
tags:
- mlx
- inkling2_loop
- code
- craft
- mirror
---

# rao-studios/Inkling-2-Mini

Rao Studios' mirror of [thinkingmachines/Inkling-2-Mini](https://huggingface.co/thinkingmachines/Inkling-2-Mini) at commit `ed40e7da625f594fc9c4990f5f66be24873c22d7`. The weights,
tokenizer, chat template and configuration are byte-for-byte that repository's; nothing was re-quantized or edited.
This card and `LICENSE` (the Apache License 2.0 text from https://www.apache.org/licenses/LICENSE-2.0.txt: [thinkingmachines/Inkling-2-Mini](https://huggingface.co/thinkingmachines/Inkling-2-Mini) declares `license: apache-2.0` in its card at `ed40e7da625f594fc9c4990f5f66be24873c22d7` and ships no LICENSE file) were added.

It exists so every install of Craft fetches the same bytes from a repository Rao controls: Craft's onboarding recommends it by this Mac's memory, and pins this repository at one commit.

## Craft

[Craft](https://craft.rao.nyc) is a coding app that runs on your Mac: the model and the index of your project stay on the
machine. Craft offers it on Macs with 48 GB of memory or more, and recommends it from 48 GB up.

Craft builds a whole iPhone app from an empty folder in one turn. The model plans, then calls fifteen agents from Craft's
store in order: the project, the look, the mark, the typed client, the mock world, each screen's kit, the navigation, the
tests and the screenshots. Then it builds, tests and shoots. Measured twice, the app built, its tests passed, and all 62 of
the original app's screens matched it, light and dark: the quality Claude Opus reached on the same app.

Those runs used Qwopus 3.5 122B on an M4 Max with 128 GB (2026-11-16), not this model. In Craft, Settings › Guide walks
through the run, says which model and Mac it needs, and sets Craft up for it. More at [craft.rao.nyc](https://craft.rao.nyc).

## Lineage

- [thinkingmachines/Inkling-2-Mini](https://huggingface.co/thinkingmachines/Inkling-2-Mini), by Thinking Machines Lab (Apache-2.0).
- [thinkingmachines/Inkling-2-Mini](https://huggingface.co/thinkingmachines/Inkling-2-Mini): the published checkpoint itself (NVFP4 and MXFP8 experts, BF16 elsewhere), which MLX reads as it is — the files here.

## Shape (from `config.json`)

| | |
|---|---|
| Architecture | `inkling2_loop`, 24 layers |
| Quantization | NVFP4, as published (NVIDIA ModelOpt; the tensors say each layer's format) |
| Weights | 19.5 GB of safetensors, 1,244 tensors |
| Context | 1,048,576 tokens |
| Tools | Inkling frames: the chat template declares every tool's JSON schema in one system frame, and the model answers with a frame per call: the tool's name, then a JSON object of the name and its `args` |

A call, as the model writes it:

```
<|message_model|>edit_file<|content_invoke_tool_json|>{"name":"edit_file","args":{"new_string":"print(\"hello\")","old_string":"print(\"hi\")","path":"Sources/App/main.swift"}}<|end_message|>
```

## Safety pass

Run by Craft's `Scripts/release/mirror-model.py` on 2026-11-20, before upload:

- Every file is one an MLX model needs; no pickle format and no native code was in the source.
- Every weight file is safetensors, its header read by hand: known dtypes, tensors that tile the
  file exactly, the same tensor names as the index.
- Every file matches the SHA-256 the Hub recorded for it at `ed40e7da625f594fc9c4990f5f66be24873c22d7`.
- No `auto_map` (remote code) in the config or the tokenizer config; a chat template is present.
- `hf_quant_config.json` was dropped: the quantizer's per-layer notes (NVIDIA ModelOpt): the tensors' own shapes and dtypes say each layer's format.
- `mtp.safetensors` was dropped: the multi-token-prediction head, for speculative decoding: Craft's engine decodes one token at a time.
- `processor_config.json` was dropped: the multimodal processor's settings: Craft reads text only.
- `tiktoken/tokenizer.model` was dropped: the vocabulary again, in tiktoken's format: Craft reads tokenizer.json.

| File | Bytes | SHA-256 |
|---|---:|---|
| `chat_template.jinja` | 6,868 | `53a9035f3255945704b5857657cfb00c1165090c0641c4f782a0cbf268cc16dc` |
| `config.json` | 2,722 | `a7a349601f0984c4abf999c480ef775d8567d21f5f84b4e8a3e035b44f288dc2` |
| `generation_config.json` | 98 | `a08737c07a65a8fdf69f7cf086576210e5aad4b241437090b26d7b0c4f7905f7` |
| `model-00001-of-00001.safetensors` | 19,526,019,168 | `65e60f51aa1284d4fb022584908c315270334d991828bb774af60584523539d1` |
| `model.safetensors.index.json` | 109,300 | `c1631065b7b00787993522dddc66a68b3abee869192977fb3dcb796e63bc5838` |
| `special_tokens_map.json` | 529 | `c16f44b3594563d82743ddb0417a2be1bfe3e2000dfad2da4b8f32554593e1a0` |
| `tokenizer.json` | 26,178,678 | `7f332727f9a3f093dd452a39cc978e6757627e5dad1b0673aa4e75174f236dd6` |
| `tokenizer_config.json` | 13,235 | `bf09342d285139fd37ce10c70878af1d52cb390e9901e10248654293a80ee636` |

## Use with mlx

Craft loads it through Frigate, which reads `inkling2_loop` checkpoints as published; mlx-lm reads it from the version that adds the model type.