TML-d24-varloop-research-GGUF Vision-Language Model: 24GB+ VRAM

November 22, 2026

TML-d24-varloop-research-GGUF Vision-Language Model: 24GB+ VRAM

At a Glance

Item Value
Repository ggml-org/TML-d24-varloop-research-GGUF
Published 2026-11-20
Formats GGUF
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

What we checked ourselves

  • It needs about 17% more tokens than the Qwen3 tokenizer for the same Japanese text, so only about 0.85× as much Japanese fits in the same context length.

Details and conditions are in “Our Own Measurements” below.

Overview

ggml-org has released “ggml-org/TML-d24-varloop-research-GGUF", which is a GGUF quantized version of the research multimodal model developed by Thinking Machines Lab. This model is based on the original model thinkingmachines/TML-d24-varloop-research and supports both text and image inputs. It adopts an architecture that allows changing the number of loop layer passes during inference, and is provided as a research model to verify how additional computational load affects internal processing and capabilities.

Specifications

  • Parameters: Total parameters 26.4B, active parameters 3.4B (at 1 loop pass; each additional loop pass increases effective active parameters by approximately 1.0B per token)
  • Architecture: 24-layer Mixture-of-Experts (MoE). 8 out of 128 total experts are routed per token. Layers 6 to 17 operate as a variable loop core
  • Context Length: Up to 1M tokens

Performance

Performance data for this model is published in the model card of the original model, “thinkingmachines/TML-d24-varloop-research". The table below shows the benchmark scores for each loop count (1 to 5 passes) in the original model and the measurement results of the comparison target “Inkling 2 Mini (3 passes)" (some columns from the original table have been omitted for readability).

→ Scroll horizontally to see all columns

Inkling 2 Mini / (3 passes) TML-d24-varloop-research / 1 pass TML-d24-varloop-research / 3 passes TML-d24-varloop-research / 5 passes
Agentic Coding
DeepSWE v1.1 mini-swe-agent v2.4.6 48.2% 0.7% 47.4% 48.2%
SWEBench Pro v2 (Hard) mini-swe-agent v2.4.6 77.9% 39.7% 78.9% 78.4%
SWEBench Multilingual bash only 73.7% 40.8% 71.8% 72.0%
SWEBench Verified bash only 79.2% 56.3% 78.7% 79.6%
Agentic Search, Tool Use, and Knowledge Work
BrowseComp original, no context management 76.0% 29.4% 65.7% 71.0%
MCP Atlas Public 82.7% 70.0% 79.4% 80.0%
JobBench 49.6% 22.8% 63.8% 63.3%
EnterpriseOps-Gym LangChain ReAct loop 36.7% 25.7% 36.6% 37.6%
Vision Interpretation and Creation
GDP.pdf 16.8% 4.6% 19.6% 20.0%
PerceptionBench 57.5% 45.5% 53.9% 52.7%
MMMU Pro Standard 10 74.9% 63.6% 71.4% 73.2%
OmniDocBench v1.6 92.8% 86.6% 92.1% 92.5%
BabyVision original 53.4% 32.7% 50.3% 49.2%
BabyVision with Python 88.8% 67.2% 88.9% 89.5%
ScreenSpot Pro original 76.5% 54.4% 72.0% 73.1%
ScreenSpot Pro with Python 87.7% 80.7% 87.7% 88.1%
Charxiv RQ original 80.6% 75.0% 79.9% 80.4%
Charxiv RQ with Python 88.3% 84.7% 87.6% 87.9%
Chartography original 22.1% 14.5% 21.6% 22.6%
Chartography with Python 72.5% 49.1% 72.3% 72.3%
ZeroBench (main) original 17.0% 9.0% 19.0% 17.0%
ZeroBench (main) with Python 34.0% 16.0% 40.0% 37.0%
Humanity’s Sixth Sense 23.8% 22.4% 25.5% 27.0%
MathVision 89.0% 74.8% 86.6% 88.1%
MuirBench 77.5% 73.6% 77.1% 77.8%
BVB 68.4% 53.9% 68.5% 69.5%
Reasoning, Knowledge, and Epistemics
HLE-Diamond 15.2% 6.3% 9.5% 11.1%
HLE (text only) 30.7% 10.5% 24.9% 25.9%
GPQA Diamond 86.6% 69.6% 84.2% 83.1%
Apex 2025 34.4% 3.1% 19.8% 18.8%
ComplexConstraints v1.1 openhands v1.45.0 33.4% 4.5% 37.9% 38.9%
SimpleQA Verified 8.4% 6.3% 7.3% 6.7%
AA-Omniscience v1.1 (Public) accuracy 21.3% 15.5% 18.4% 17.4%
AA-Omniscience v1.1 (Public) non-hallucination rate 64.6% 79.3% 84.0% 85.9%
SimpleVQA 45.4% 39.0% 42.4% 41.3%
WorldVQA 15.2% 11.3% 14.2% 13.5%
GraphWalks BFS 256K 69.7% 23.5% 64.3% 63.8%
MRCR v2 8-needle 256K–512K 46.4% 27.2% 89.2% 92.1%
Safety and Alignment
FORTRESS (adversarial) Public 89.5% 89.0% 90.0% 89.4%
FORTRESS (benign) Public 96.4% 95.0% 95.3% 95.0%
StrongREJECT (none) 97.8% 98.8% 98.2% 98.4%

According to the publisher’s measurement results, the model exhibits a characteristic where increasing the number of loop passes significantly improves scores in complex domains such as coding and reasoning. On “SWEBench Verified bash only", which measures the ability to fix real GitHub issues, accuracy rises from 56.3% with 1 pass to 78.7% with 3 passes and 79.6% with 5 passes, matching the high accuracy of Inkling 2 Mini (79.2%). Additionally, on “GPQA Diamond", which handles advanced PhD-level problems, the Pass@1 accuracy jumps from 69.6% with 1 pass to 84.2% with 3 passes. On the other hand, safety and alignment metrics such as “FORTRESS (adversarial) Public" show no significant changes, recording 89.0% for 1 pass, 90.0% for 3 passes, and 89.4% for 5 passes. Thus, while increasing loop computations drastically enhances performance on complex thinking tasks, behavior in areas like safety maintains stable scores.

Strengths and Use Cases

This model is a conversational multimodal model that accepts both text and image inputs. It is provided as a research model to evaluate changes in inference compute and internal model behavior within a variable loop structure. Its strengths and intended use cases are based on the characteristics of the original model “thinkingmachines/TML-d24-varloop-research".

Adjusting Inference Compute According to Tasks

Because the number of loop layer passes can be modified during inference, computational load can be adjusted to match task difficulty. For coding tasks involving automatic code correction and repository operations, and reasoning tasks requiring complex logical flow, increasing the loop count can extract higher processing accuracy.

Integrated Understanding and Dialogue of Images and Text

It supports pixel-based image input (recommended resolution: 40px to 4096px per side). Visual information inputted via the Vision Transformer is projected into a shared latent space along with text and processed in an integrated manner. It can be used for dialogues involving visual elements such as document image reading, chart analysis, and screen state comprehension.

Parameter Control of Reasoning Depth (reasoning_effort)

It supports the reasoning_effort parameter for setting the model’s thinking depth. You can specify string labels such as none, minimal, low, medium, high (default), xhigh, and max, or a numerical value from 0.0 to 0.99 to adjust inference depth.

Note that this model is positioned for research purposes and has not been tuned for production deployment. For general practical use, utilizing production-grade models (such as Inkling 2 Mini) is recommended, and this model is primarily designed for investigating and experimenting with loop computation structures during inference.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
24GB (RTX 4090 / 3090, etc.) IQ4_NL 16.4GB 19.7GB
32GB (RTX 5090, etc.) Q8_0 25.9GB 31.1GB
80GB class (A100 / H100) BF16 48.6GB 58.4GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp as-is.

It is distributed in GGUF, so no conversion is needed.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists.

Our Own Measurements

Values we checked ourselves on our server (no GPU) by actually reading this model’s files, without running the model — not figures copied from the model card. How we measure, and the results for every model: Our Measurements.

Measurements We Did Not Take

This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation). Only values we checked without running the model, such as token counts and the GGUF header, are shown.

Japanese Token Efficiency

Tokenizer Tokens per 1,000 Japanese characters Ratio to the same text in English
This model 806 1.38×
Qwen3 688 1.26×
Llama 3.2 744 1.36×
Gemma 3 564 1.03×
gpt-oss 795 1.45×
LLM-jp-3 497 0.85×

It needs about 17% more tokens than the Qwen3 tokenizer for the same Japanese text, so only about 0.85× as much Japanese fits in the same context length.

Counted with the tokenizer.json of thinkingmachines/TML-d24-varloop-research on a fixed text we wrote ourselves (876 Japanese characters across news, conversation, technical docs, a formal email, travel writing and a recipe) and its English translation. Fewer tokens mean more Japanese fits in the context window.

Inside the GGUF File

File: TML-d24-varloop-research-IQ4_NL.gguf (16.42GB, IQ4_NL). We read only the header (metadata) of the file, not the weights.

Item Value
Architecture (as named in the GGUF) inkling2_loop
Maximum trained context length 1,048,576 tokens
Layers 24
Experts 8 active out of 128
Vocabulary size 200,058
Chat template Included (mentions tool calls, has a thinking switch)
imatrix Used (inkling2-loop-calibration-v6.txt, 1,770 chunks)
Weight types (share of parameters) IQ4_NL 78.2% / Q8_0 21.7% / F32 0.1%
Average bits per weight 5.41 bits
Embedding / output layer type Q8_0 / Q8_0

The quant name in the file name describes the file as a whole; in practice layers mix several types. The average bits per weight is the measured file data divided by the number of weights.

How to Get It

This model, “ggml-org/TML-d24-varloop-research-GGUF", is available in GGUF format on the Hugging Face repository. Gated license agreement procedures for downloading are not required.

Compatible engines such as llama.cpp or llama.app can be used as execution environments. To launch an inference server, run the following command:

llama serve -hf ggml-org/TML-d24-varloop-research-GGUF:IQ4_NL

If you want to change the number of loop layer passes to an arbitrary value (e.g., 2 passes) when starting the server, add the --override-kv option as follows:

llama serve -hf ggml-org/TML-d24-varloop-research-GGUF:IQ4_NL --override-kv inkling2_loop.loop.count=int:2

The distribution formats include IQ4_NL calibrated with imatrix, as well as Q8_0 and BF16. In the IQ4_NL quantized version, the expert weights of layers 0 to 19 are quantized to IQ4_NL, while the F32 weights of the loop controller and other weight matrices are maintained at Q8_0 precision.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-22: Added our own measurements: Japanese token efficiency, what is inside the GGUF file.