Xylolabs

TML-d24-varloop-research MLX

Base model Format License Code

thinkingmachines/TML-d24-varloop-research as MLX weights for mlx-lm: routed experts in the original NVFP4 (layers 0 to 19, group size 16, with the per-expert scales) and MXFP8 (layers 20 to 23, group size 32) exactly as shipped; attention, embedding and LM head in affine 8-bit (group size 64); vision tower, norms and the loop controllers in bf16 or f32. The multi-token-prediction head of the original is not included.

How to run

mlx-lm 0.32.2 does not know this architecture. Install the pull request that adds it (mlx-lm PR 2086), the commit this build was measured with:

pip install -U "git+https://github.com/ml-explore/mlx-lm.git@0c4c99192230f1943dd468d297f0b8a64ae7b7c2"

mlx_lm.generate --model Xylolabs/TML-d24-varloop-research-MLX --prompt 'Explain how a heat pump works.'
mlx_lm.chat     --model Xylolabs/TML-d24-varloop-research-MLX

The model runs its looped layers (6 to 17) for the 3 passes set in config.json. Image input is part of the architecture and the vision tower is included; image input has not been checked.

Accuracy

Mean KL divergence from BF16 (lower is better) and top-1 agreement with BF16.

Quant Size WikiText-2 KL top-1 Chat KL top-1 Held-out KL top-1
MLX: experts NVFP4 and MXFP8, attention, embedding and LM head 8-bit 18.09 GB 0.003331 97.436% 0.019627 96.696% 0.002881 97.520%
Mean KL divergence by file size Top-1 agreement by file size

Which file for your hardware

Macs only (MLX runs on Apple silicon; memory options as Apple lists them for the Mac Studio, M5 Max 36, 48, 64 and 128 GB, M5 Ultra 96, 256 and 512 GB, and the 14-inch MacBook Pro M5 Max, checked 2026-11-22). The GPU may use about 75% of unified memory by default; sudo sysctl iogpu.wired_limit_mb=<MB> raises that limit (it resets at reboot). Fit rule: file + KV cache at the stated context + about 2 GB must fit in the GPU limit; sizes in GB (10^9 bytes); context is the longest that fits at the default limit, at least 32K. KV cache 33 KB per token for the 8 full-attention cache slots (layers 11 and 17 are visited three times by the loop), plus a fixed 170 MB for the 40 sliding-window slots. No MLX speed has been measured on these Macs.

Hardware Memory Recommended file (ours only) Context that fits Notes
Mac Studio M5 Max (also MacBook Pro 14-inch M5 Max, 32-core GPU) 36 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 200K Fits at the default GPU limit (75% of memory). Raising it to 90% (sudo sysctl iogpu.wired_limit_mb=33177) allows 361K.
Mac Studio M5 Max (also MacBook Pro 14-inch M5 Max, 40-core GPU) 48 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 469K Fits at the default GPU limit (75% of memory). Raising it to 90% (sudo sysctl iogpu.wired_limit_mb=44236) allows 683K.
Mac Studio M5 Max (also MacBook Pro 14-inch M5 Max, 40-core GPU) 64 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 826K Fits at the default GPU limit (75% of memory). Raising it to 90% (sudo sysctl iogpu.wired_limit_mb=58982) allows 1024K.
Mac Studio M5 Ultra 96 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 1024K (model maximum) Fits at the default GPU limit (75% of memory).
Mac Studio M5 Max, 40-core GPU (also MacBook Pro 14-inch) 128 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 1024K (model maximum) Fits at the default GPU limit (75% of memory).
Mac Studio M5 Ultra 256 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 1024K (model maximum) Fits at the default GPU limit (75% of memory).
Mac Studio M5 Ultra, 80-core GPU 512 GB unified Xylolabs/TML-d24-varloop-research-MLX (NVFP4/MXFP8 experts, 8-bit rest, 18.09 GB) 1024K (model maximum) Fits at the default GPU limit (75% of memory).

License and notices

These files are under the Apache License 2.0, copied in LICENSE; use is subject to Thinking Machines' Model Acceptable Use Policy. We converted the weights to MLX with mlx-lm (experts unchanged, other linear layers in affine 8-bit) and left out the multi-token-prediction head. The source revision and changes are in ORIGINS.md and NOTICE. Not affiliated with or endorsed by Thinking Machines Lab.

Updates

  • 2026-11-22: Added a hardware recommendation table.
  • 2026-11-22: First release. Checked with mlx-lm pull request 2086 (commit 0c4c99192230f1943dd468d297f0b8a64ae7b7c2); mlx-lm 0.32.2 does not include the Inkling 2 model.

Thinking Machines' original model description

Below is the model card of thinkingmachines/TML-d24-varloop-research (Apache-2.0), without its metadata header and unchanged.

TML-d24-varloop-research

1. General Information

TML-d24-varloop-research is a multimodal research model that accepts text and image inputs and generates text outputs. It is a variable-looped counterpart to Inkling 2 Mini: whereas Inkling 2 Mini always runs its looped layers a fixed number of times, the number of loops in this model can be chosen at inference time. We are releasing it with open weights to support research on looped models. Because the number of loops can be varied on a single set of weights, the model allows cleaner comparisons of how additional looped computation affects a model's internal computation and overall capabilities.

Languages: English, general multilingual support

Model provider: Thinking Machines Lab, Inc.

Date of release: Nov 20, 2026

License: Apache 2.0

Model Acceptable Use Policy

2. Getting Started

The weights are available for download through Hugging Face here.

The model is not available through an API, including Tinker. It is distributed as open weights only.

Required hardware

The model is distributed as a single checkpoint of about 20GB. Its expert layers are NVFP4, the last four use MXFP8, and the remaining weights are BF16.

The checkpoint requires a single machine with at least 32GB of GPU or unified memory. This can be met with any of the following example configurations:

  • 1x NVIDIA DGX Spark
  • 1x NVIDIA GPU with 32GB or more of VRAM (for example, GeForce RTX 5090 or RTX PRO 6000 Blackwell)
  • Apple silicon Mac with 32GB or more of unified memory (for example, Mac mini)

The model uses the same mixed NVFP4 and MXFP8 precision it was trained in, so this checkpoint is its native precision and needs no further quantization.

Required software

TML-d24-varloop-research supports local inference using the following frameworks, along with their dependency libraries:

TML-d24-varloop-research 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 a variable number of times with shared weights, chosen at inference time; the model was trained with 1 to 5 passes. 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 per token. This corresponds to 3.4B active parameters at 1 pass, 5.4B at 3 passes, and 7.4B at 5 passes.

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.

TML-d24-varloop-research 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.

Training with variable looping

TML-d24-varloop-research shares its pretraining and first stage of midtraining with Inkling 2 Mini. For the rest of midtraining, it was trained with 1 to 5 passes through the looped layers, sampled in a 1:2:6:2:1 ratio, and it was then post-trained with 3 passes. Because 3 passes dominated training once looping was introduced, performance at other pass counts may understate what models trained specifically at those counts could achieve. The two models' later training stages differ, so we caution against comparing TML-d24-varloop-research directly to Inkling 2 Mini.

5. Evaluations

Each row shows one benchmark score for TML-d24-varloop-research run at 1 to 5 passes through its looped layers, beside Inkling 2 Mini, which always runs 3 passes. Scores are in percent; – marks a pass count without a completed result yet.

Inkling 2 MiniTML-d24-varloop-research
(3 passes)1 pass2 passes3 passes4 passes5 passes
Agentic Coding
DeepSWE v1.1
mini-swe-agent v2.4.6
48.2%0.7%28.3%47.4%46.9%48.2%
SWEBench Pro v2 (Hard)
mini-swe-agent v2.4.6
77.9%39.7%69.1%78.9%76.0%78.4%
SWEBench Multilingual
bash only
73.7%40.8%68.5%71.8%72.8%72.0%
SWEBench Verified
bash only
79.2%56.3%76.5%78.7%79.3%79.6%
Agentic Search, Tool Use, and Knowledge Work
BrowseComp
original, no context management
76.0%29.4%53.8%65.7%69.4%71.0%
MCP Atlas
Public
82.7%70.0%77.6%79.4%79.7%80.0%
JobBench49.6%22.8%55.1%63.8%61.8%63.3%
EnterpriseOps-Gym
LangChain ReAct loop
36.7%25.7%34.4%36.6%36.5%37.6%
Vision Interpretation and Creation
GDP.pdf16.8%4.6%11.2%19.6%18.0%20.0%
PerceptionBench57.5%45.5%51.3%53.9%53.7%52.7%
MMMU Pro
Standard 10
74.9%63.6%69.8%71.4%72.1%73.2%
OmniDocBench v1.692.8%86.6%90.5%92.1%91.8%92.5%
BabyVision
original
53.4%32.7%44.9%50.3%50.3%49.2%
BabyVision
with Python
88.8%67.2%84.1%88.9%88.8%89.5%
ScreenSpot Pro
original
76.5%54.4%66.0%72.0%73.4%73.1%
ScreenSpot Pro
with Python
87.7%80.7%86.5%87.7%87.2%88.1%
Charxiv RQ
original
80.6%75.0%78.3%79.9%79.2%80.4%
Charxiv RQ
with Python
88.3%84.7%87.7%87.6%89.3%87.9%
Chartography
original
22.1%14.5%17.8%21.6%21.4%22.6%
Chartography
with Python
72.5%49.1%69.1%72.3%73.7%72.3%
ZeroBench (main)
original
17.0%9.0%13.0%19.0%18.0%17.0%
ZeroBench (main)
with Python
34.0%16.0%31.0%40.0%35.0%37.0%
Humanity’s Sixth Sense23.8%22.4%24.8%25.5%26.5%27.0%
MathVision89.0%74.8%84.9%86.6%87.8%88.1%
MuirBench77.5%73.6%76.1%77.1%76.4%77.8%
BVB68.4%53.9%64.4%68.5%69.2%69.5%
Reasoning, Knowledge, and Epistemics
HLE-Diamond15.2%6.3%8.5%9.5%9.7%11.1%
HLE (text only)30.7%10.5%20.0%24.9%25.4%25.9%
GPQA Diamond86.6%69.6%79.2%84.2%81.9%83.1%
Apex 202534.4%3.1%10.4%19.8%15.6%18.8%
ComplexConstraints v1.1
openhands v1.45.0
33.4%4.5%26.0%37.9%39.6%38.9%
SimpleQA Verified8.4%6.3%6.5%7.3%6.8%6.7%
AA-Omniscience v1.1 (Public)
accuracy
21.3%15.5%16.3%18.4%18.0%17.4%
AA-Omniscience v1.1 (Public)
non-hallucination rate
64.6%79.3%85.0%84.0%85.7%85.9%
SimpleVQA45.4%39.0%41.8%42.4%42.9%41.3%
WorldVQA15.2%11.3%13.0%14.2%14.3%13.5%
GraphWalks BFS
256K
69.7%23.5%56.0%64.3%64.3%63.8%
MRCR v2
8-needle 256K–512K
46.4%27.2%65.2%89.2%93.4%92.1%
Safety and Alignment
FORTRESS (adversarial)
Public
89.5%89.0%90.4%90.0%89.9%89.4%
FORTRESS (benign)
Public
96.4%95.0%95.9%95.3%95.7%95.0%
StrongREJECT (none)97.8%98.8%98.4%98.2%99.4%98.4%

6. Safety

We evaluated this model with the same pre-release process as Inkling 2 Mini. It is less capable than models whose weights we have already released on CBRN and cyber evaluations, and it refuses harmful requests about as reliably as Inkling 2 Small.

For production use, we recommend Inkling 2 Mini as it generally performs better on capabilities and alignment evaluations. We do not intend TML-d24-varloop-research for production use.

Across all areas, we concluded that TML-d24-varloop-research did not present risk of material uplift beyond what’s already available in the open-weight ecosystem.

7. Bias, risks and limitations

TML-d24-varloop-research is a research model, not a production model. It 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 large-scale models trained on web-derived and synthetic data, TML-d24-varloop-research 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.

TML-d24-varloop-research’s knowledge is limited to information available as of its training cutoff, and it may not reflect events, developments, or changes that occurred afterward.

TML-d24-varloop-research is intended to be a research artifact. It has not been carefully tuned for deployment, and may behave unpredictably at pass counts other than 3. We do not recommend deploying this model in production or user-facing applications.

8. Legal

Training Data Documentation

Downloads last month
8
Safetensors
Model size
26B params
Tensor type
U32
·
F32
·
BF16
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

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

Model tree for Xylolabs/TML-d24-varloop-research-MLX

Quantized
(4)
this model

Collection including Xylolabs/TML-d24-varloop-research-MLX