Inkling-2-Mini / README.md
rpakala's picture
Mirror thinkingmachines/Inkling-2-Mini@ed40e7da625f, after Craft's safety pass
82141c1 verified
|
Raw History Blame Contribute Delete
5.13 kB
metadata
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 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 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 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.

Lineage

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.