Paris99's picture
Card title: + DSpark
0a45a05 verified
|
Raw History Blame Contribute Delete
8.66 kB
metadata
license: apache-2.0
license_link: https://www.apache.org/licenses/LICENSE-2.0
base_model: thinkingmachines/Inkling-2-Mini
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: mlx
tags:
  - mlx
  - apple-silicon
  - inkling2_loop
  - moe
  - dspark
  - speculative-decoding

Inkling 2 Mini for Apple Silicon (MLX, 4/8-bit, lossless, + DSpark)

thinkingmachines/Inkling-2-Mini (revision 6e7c824) for MLX on Apple Silicon, with its DSpark speculative-decoding draft bundled (mtp.safetensors) and its vision encoder included. The routed experts are kept exactly as Thinking Machines released them: 4-bit NVFP4 in layers 0–19 and 8-bit MXFP8 in layers 20–23, repacked into MLX's native nvfp4 / mxfp8 formats with every value unchanged. Attention, embeddings, the output layer and the vision encoder stay in the original bf16. This is the release itself, in MLX. Size: 19.5 GB plus the 0.64 GB draft.

Inkling 2 Mini is Thinking Machines' looped mixture-of-experts model: 26.4B parameters, about 3.4B active per token, 24 layers of which the middle twelve run three times (48 layer passes per token), text and image input, up to 1M tokens of context.

Builds

Build Size Same next token as Thinking Machines' code KL vs their code GSM8K (50 questions) Peak memory, 8K-token prompt
netprojects/Inkling-2-Mini-MLX-4-8bit (recommended) 19.5 GB + 0.64 GB draft 98.6% 0.0018 47/50 27.8 GB
netprojects/Inkling-2-Mini-MLX-3bit 14.1 GB + 0.64 GB draft 95.7% 0.0421 48/50 21.2 GB
netprojects/Inkling-2-Mini-MLX-2.5bit 10.8 GB + 0.64 GB draft 87.3% 0.1904 45/50 18.2 GB
netprojects/Inkling-2-Mini-MLX-4bit 16.5 GB + 0.64 GB draft 95.9% 0.0286 47/50 23.5 GB

"Thinking Machines' code" is their Transformers implementation of Inkling 2 Mini, run on the released weights; the columns are explained under "What was checked".

Install

Inkling 2 Mini support is not in a released MLX-VLM yet. Install MLX-VLM with it from internet-projects/inkling2-mlx (MLX-VLM plus Inkling 2 Mini support; Python 3.10 or newer):

pip install "git+https://github.com/internet-projects/inkling2-mlx@v0.1.0" torch torchvision

PyTorch and torchvision are only used on the CPU by the image processor.

Use

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("netprojects/Inkling-2-Mini-MLX-4-8bit")
prompt = apply_chat_template(processor, model.config, "Explain in two sentences why the sky is blue.",
                             reasoning_effort="low")
print(generate(model, processor, prompt, max_tokens=2048, temperature=0.0).text)

reasoning_effort is part of Inkling's chat template: none, minimal, low, medium, high (the default), xhigh, max, or a number from 0.0 to 0.99. Thinking Machines notes that the model can think at length even at none and can get stuck in long reasoning loops, so set max_tokens.

With an image:

prompt = apply_chat_template(processor, model.config, "What error does this terminal show?", num_images=1, reasoning_effort="low")
print(generate(model, processor, prompt, image=["screenshot.png"], max_tokens=2048).text)

With the DSpark draft (greedy output stays the model's own, up to near-ties):

from mlx_vlm.speculative.drafters import load_drafter
draft, kind = load_drafter("netprojects/Inkling-2-Mini-MLX-4-8bit")
print(generate(model, processor, prompt, max_tokens=2048, draft_model=draft, draft_kind=kind,
               draft_block_size=3).text)

Server: mlx_vlm.server --model netprojects/Inkling-2-Mini-MLX-4-8bit --draft-model netprojects/Inkling-2-Mini-MLX-4-8bit --draft-kind dflash --draft-block-size 3.

OpenAI-compatible server, with tool calls and the thinking returned separately (reasoning_content):

mlx_vlm.server --model netprojects/Inkling-2-Mini-MLX-4-8bit --port 8000

Speed

MacBook Pro M5 Max (128 GB), idle, 21 November 2026; median of 3 runs. Decode: 5 prompts (math, code, chat) × 256 tokens, greedy, reasoning none. DSpark: the bundled draft at its best block size.

Build Prompt processing, 1K / 8K / 32K tokens (tok/s) Generation (tok/s) Generation with DSpark (tok/s)
4-8bit (this build) 1752 / 1658 / 1208 36.6 47.1 (×1.29, block 3)
4bit 1535 / 1412 / 1104 43.6 70.9 (×1.63, block 4)
3bit 1386 / 1351 / 1075 44.4 75.2 (×1.69, block 4)
2.5bit 1482 / 1384 / 1100 45.7 83.9 (×1.84, block 4)

The draft's default block size (8) is slower on this model; use 3 or 4 as below.

Thinking Machines' results

From the Inkling 2 Mini card, for the original model (their harnesses and settings; a selection of their table):

Benchmark (Thinking Machines' figures) Inkling 2 Mini Qwen3.8-27B Gemma 4 31B
Artificial Analysis Intelligence Index v5.0 20.1 20.1 4.4
SWE-bench Verified (bash only) 79.2% 77.8% 46.1%
SWE-bench Pro v2 Hard (mini-swe-agent) 77.9% 73.8% 14.2%
Terminal-Bench 2.1 (pi) 71.6% 73.0% 43.4%
BrowseComp (with compaction) 79.0% 61.6% 37.0%
MCP Atlas 82.7% 77.3% 54.2%
MMMU Pro 74.9% 76.3% 73.4%
OmniDocBench v1.6 92.8% 91.5% 82.5%
LongBench v2 66.9% 52.3% 52.9%
ARC-AGI v1 / v2 84.5 / 49.3% 87.5 / 42.4% –

What was checked

On a MacBook Pro M5 Max (128 GB), 21 November 2026, with MLX 0.32.4 and inkling2-mlx v0.1.0.

  • Against Thinking Machines' code. Their Transformers implementation (pull request #51050) ran on the CPU with the released weights, on 12 short chat texts (488 answer positions). This build gave the same next token at 98.6% of positions (mean KL 0.0018), and its greedy answers to 3 prompts matched theirs token for token (24 tokens each). Every tensor of this build is bit-identical to the release as MLX reads it.
  • GSM8K, 50 fixed test questions (reasoning none, greedy): 47/50.
  • Tool calling (reasoning none): an agent prompt with bash and read tools produced a well-formed call in Inkling's format, which mlx_vlm.server returns as OpenAI tool_calls.
  • Coding: 3/3 short Python tasks passed their asserts. Thinking (reasoning low): 2/2 questions answered correctly after one closed thinking block. Vision: named the red square and blue circle of a test image.
  • Memory: peak 26.9 GB with a 4K-token prompt and 27.8 GB with an 8K-token prompt (mlx-vlm defaults). With a smaller prefill step the peak drops: 22.3 / 22.6 GB at prefill_step_size=512 and 21.4 / 21.6 GB at prefill_step_size=256 (4K / 8K-token prompts, plus 0.64 GB if the DSpark draft is loaded), which fits the default GPU memory limit of a 36 GB Mac: pass prefill_step_size=256 to generate, or --prefill-step-size 256 to mlx_vlm.server.
  • DSpark: the bundled draft loads with this build; greedy output follows the model's own except at near-ties (checked: divergences only where the top two tokens were within 0.2 logits). Speeds are under "Speed".

Precision note: Thinking Machines released the routed experts in 4-bit NVFP4 (8-bit MXFP8 in the last four layers) and no higher-precision weights; MLX builds of this model keep the experts at that precision, so this 4/8-bit build already carries the full released precision.

Known limitation of the current server: when the model calls a tool, content can also carry a short sentence of its reasoning and the tool name.

License and credits

Apache-2.0, as Inkling 2 Mini. Use of the model is also governed by Thinking Machines' Model Acceptable Use Policy. LICENSE is the Apache-2.0 text and NOTICE describes the conversion. Inkling 2 Mini is by Thinking Machines Lab; please cite their work.

Published by NetProjects (Internet Projects Ltd). Not affiliated with Thinking Machines.

More MLX builds from NetProjects: huggingface.co/netprojects · netprojects.com