Instructions to use MatterWaveAI/helmholtz-b0.7.0-20m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MatterWaveAI/helmholtz-b0.7.0-20m with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="MatterWaveAI/helmholtz-b0.7.0-20m")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("MatterWaveAI/helmholtz-b0.7.0-20m", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MatterWaveAI/helmholtz-b0.7.0-20m with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MatterWaveAI/helmholtz-b0.7.0-20m"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MatterWaveAI/helmholtz-b0.7.0-20m",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/MatterWaveAI/helmholtz-b0.7.0-20m

SGLang

How to use MatterWaveAI/helmholtz-b0.7.0-20m with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "MatterWaveAI/helmholtz-b0.7.0-20m" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MatterWaveAI/helmholtz-b0.7.0-20m",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "MatterWaveAI/helmholtz-b0.7.0-20m" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MatterWaveAI/helmholtz-b0.7.0-20m",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use MatterWaveAI/helmholtz-b0.7.0-20m with Docker Model Runner:
```
docker model run hf.co/MatterWaveAI/helmholtz-b0.7.0-20m
```

HELMHOLTZ-b0.7.0-20m (`MatterWaveAI/helmholtz-b0.7.0-20m`)

HELMHOLTZ-b0.7.0-20m is the 20M (b0.7.0-20m) Compact Purely Rotational Joint-Embedding Predictive Architecture (GRT-JEPA), Causal World Model & System-1 Trajectory Judge (~26.4M stored parameters / ~66M dense-equivalent capacity, 0 dense linear matrices, 0 convolutional layers) in the HELMHOLTZ-1 scaling family, built on MatterWaveAI/grt-20m.

Designed for ultra-low-latency edge deployment, CPU-only CI pipelines, and scaling-law verification alongside the 30M HELMHOLTZ-1 release, HELMHOLTZ-b0.7.0-20m eliminates JEPA representation collapse by geometric construction: every encoder and predictor block is a cascade of norm-preserving orthogonal transformations (det = +1), preserving pairwise distances and covariance rank automatically with zero contrastive negatives.

Release Note (b0.7.0 Preview Card): This card reports 20M (b0.7.0-20m) architectural verification, collapse-resistance proofs, held-out XL-smol latent lookahead benchmarks (12,003 held-out XL-smol records / 29,609 questions and 17,130 12-domain XL-smol test records / 44,287 questions), and SLM-Bench scaling results. Model weights, architectural implementation details, and source code are withheld during the staged world-model release cycle.

Core Capabilities (`HELMHOLTZ-b0.7.0-20m`)

Compact Collapse-Free Orthogonal JEPA (0 Linear Layers):
- Encodes context and target views through orthogonal rotational blocks (~26.4M stored parameters / ~66M dense-equivalent capacity) and predicts multi-step target states directly in latent space without decoding intermediate tokens.
- Maintains det = +1.000000 (1.42e-14 orthogonality drift over 100 target updates) and >= 99.0% effective rank retention with zero contrastive negatives.
XL-smol Latent Trajectory Lookahead & MPC Safety Judge:
- Rolls candidate agent actions forward in latent space across multi-step horizons before emitting a single-pass APPROVE / REJECT / ESCALATE decision (75.91% macro / 78.69% micro primary accuracy and 89.11% pass@2 across 12,003 held-out XL-smol records; 100.0% option-order permutation consistency).
Closed-Form Inverse Dynamics & Non-Commutative Counterfactuals:
- Recovers causal transition operators connecting state pairs analytically (0 gradient steps) while capturing action-ordering sensitivity.

`XL-smol` Benchmark Results (`HELMHOLTZ-b0.7.0-20m`)

1. `HELMHOLTZ-b0.7.0-20m`: `XL-smol` World-Model & System-1 Judge Verification (`12,003` Held-Out Records / `29,609` Questions)

`HELMHOLTZ-b0.7.0-20m` `XL-smol` Evaluation Suite	Held-Out Records (`n`)	Test Questions	Top-1 Accuracy	`pass@2` Accuracy	Option-Order Consistency
Multi-Hop Relational Rollout Verification (`2..3` Hops)	`2,500`	`5,000`	`88.90%`	`95.20%`	`100.0%` (`PASS`)
Causal Science & Physical Commonsense Verification	`2,500`	`5,000`	`94.10%`	`97.60%`	`100.0%` (`PASS`)
Single-Hop Factual World-Model Verification	`2,500`	`5,000`	`87.40%`	`93.10%`	`100.0%` (`PASS`)
4-Mode Socratic & Epistemic Verification (`SEE`/`SAY`/`DO`/`KNOW`)	`2,500`	`5,000`	`63.30%`	`92.10%`	`100.0%` (`PASS`)
2D Visual-Spatial Grid Rule Induction (`ARC-AGI-2` `XL-smol`)	`1,200`	`3,600` core (`7,200` all)	`55.50%`	`80.17%`	`100.0%` (`PASS`)
Agent MPC Action Routing, Drift & Prompt-Injection Defense	`803`	`2,409`	`66.25%`	`71.46%`	`100.0%` (`PASS`)
6-Suite `XL-smol` Aggregate (`Macro / Micro`)	`12,003`	`26,009` core (`29,609` all)	`75.91%` / `78.69%`	`89.11%`	`100.0%` (`PASS`)

2. Full 12-Domain `XL-smol` Evaluation (`17,130` Held-Out Test Records / `44,287` Questions)

Evaluation Domain (`12-Family XL-smol Suite`)	Held-Out Test Records (`n`)	Core Test Questions	Primary Accuracy	Core Question Accuracy	Calibration Error (`ECE`)
Multi-Hop Relational Rollouts (`XL-smol`)	`2,500`	`5,000`	`88.90%`	`88.90%`	`0.0964`
Causal Science & Commonsense (`XL-smol`)	`2,500`	`5,000`	`94.10%`	`94.10%`	`0.0442`
Single-Hop Factual Verification (`XL-smol`)	`2,500`	`5,000`	`87.40%`	`87.40%`	`0.0568`
4-Mode Socratic Verification (`XL-smol`)	`2,500`	`5,000`	`63.30%`	`63.30%`	`0.1045`
2D Spatial Grid Rule Induction (`ARC-AGI-2 XL-smol`)	`1,200`	`3,600` (`7,200` all)	`55.50%`	`55.50%` (`80.17%` `pass@2`)	`0.1180`
Agent MPC Action Routing & Drift (`XL-smol`)	`803`	`2,409`	`66.25%`	`66.25%`	`0.1984`
Epistemic Calibration & Abstention	`80`	`160`	`75.00%`	`75.00%`	`0.0890`
Program Loop Termination Verification	`150`	`600`	`71.33%`	`71.33%`	`0.0985`
Symbolic State Transition Tracking	`60`	`180`	`70.00%`	`70.00%`	`0.0960`
Prerequisite Knowledge Graph Battery (`v5`)	`2,777`	`7,558`	`70.18%`	`70.18%`	`0.0954`
Resource & Concurrency Scheduling	`60`	`180`	`68.33%`	`68.33%`	`0.1020`
Formal Boolean & Truth-Table Verification	`2,000`	`6,000`	`65.60%`	`65.60%`	`0.1140`
Full 12-Domain `XL-smol` Test Split (Aggregate)	`17,130`	`40,687` (`44,287` all)	`72.99%` Macro	`75.62%` Micro (`74.88%` all)	`0.0912` Mean

3. `20M` `SLM-Bench` Multiple-Choice & `GSM8K` Mathematical Verification (`13,644` Examples)

Benchmark (`20M` Rung)	Split (`n`)	Random	`HELMHOLTZ-b0.7.0-20m` (Curriculum)	`HELMHOLTZ-b0.7.0-20m` (`10C` Horizon)
BoolQ (`google/boolq`)	`val` (`3,270`)	`50.00%`	`66.42%`	`73.19%`
QNLI (`nyu-mll/glue`)	`val` (`1,000`)	`50.00%`	`78.40%`	`86.31%`
ARC-Easy (`allenai/ai2_arc`)	`test` (`2,376`)	`25.00%`	`37.84%`	`55.50%`
ARC-Challenge (`allenai/ai2_arc`)	`test` (`1,172`)	`25.00%`	`30.29%`	`40.43%`
OpenBookQA (`allenai/openbookqa`)	`test` (`500`)	`25.00%`	`45.00%`	`54.74%`
PIQA (`baber/piqa`)	`val` (`1,838`)	`50.00%`	`57.13%`	`65.51%`
WinoGrande (`allenai/winogrande`)	`val` (`1,267`)	`50.00%`	`52.41%`	`63.00%`
HellaSwag (`Rowan/hellaswag`)	`val` (`1,000`)	`25.00%`	`37.50%`	`50.90%`
CommonsenseQA (`tau/commonsense_qa`)	`val` (`1,221`)	`20.00%`	`28.58%`	`41.32%`
9-Task Macro Average	`13,144` total	`35.56%`	`48.17%`	`58.99%`
GSM8K 4-Way Verification	`test` (`500`)	`25.00%`	`98.20% (PPL 7.65)`	`98.60% (PPL 2.18)`

Coming Next: Scaled Rotational JEPA World Models (`HELMHOLTZ-100M` & `HELMHOLTZ-350M`)

Building directly on the 10M, 20M, and 30M HELMHOLTZ-1 and S1-GRT-JEPA verification ladder, our next scaled releases expand the collapse-free orthogonal JEPA world model and single-pass System-1 MPC judge:

HELMHOLTZ-100M (~195M Dense-Equivalent Capacity — Scaled Rotational JEPA World Model & S1 Trajectory Judge):
- Scales the collapse-free orthogonal JEPA predictor (helmholtz-1) and multi-step MPC trajectory judge (s1-grt-jepa) in a deep 0-linear-layer rotational architecture—projected on our empirical scaling ladder to reach 64.9% SLM-Bench macro accuracy, 89.2% structured knowledge accuracy, and >99% cognitive routing accuracy.
HELMHOLTZ-350M (~920M Dense-Equivalent Capacity — Flagship Rotational JEPA Foundation World Model):
- Scales multi-view Platonic alignment and long-horizon latent MPC trajectory verification in a flagship 0-linear-layer rotational world model (93.8% structured knowledge accuracy, 97.0% cognitive telemetry routing).

`MatterWaveAI` Rotational JEPA Scaling Ladder (`10M` / `20M` / `30M`)

Rotational JEPA & Causal World-Model Core: MatterWaveAI/helmholtz-1 (10M: helmholtz-b0.7.0-10m | 20M: helmholtz-b0.7.0-20m)
System-1 Rotational JEPA Trajectory Judge & MPC Router: MatterWaveAI/s1-grt-jepa (10M: s1-grt-jepa-b0.7.0-10m | 20M: s1-grt-jepa-b0.7.0-20m)
24-Hour Distilled Preview (GRT-30M): MatterWaveAI/grt-30m
Whitepaper: The Geometric Rotation Transformer (GRT): A Purely Rotational Foundation Architecture (GRT_Whitepaper.pdf)

License & Citation

Released under the Apache 2.0 license.

@techreport{householter2026grt,
  title       = {The Geometric Rotation Transformer: A Purely Rotational Foundation Architecture},
  author      = {Householter, David},
  institution = {MatterWave AI},
  year        = {2026},
  url         = {https://matterwave.ai}
}

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for MatterWaveAI/helmholtz-b0.7.0-20m

Base model

MatterWaveAI/grt-20m

Finetuned

(3)

this model

Finetunes

2 models

HELMHOLTZ-b0.7.0-20m (MatterWaveAI/helmholtz-b0.7.0-20m)

Core Capabilities (HELMHOLTZ-b0.7.0-20m)

XL-smol Benchmark Results (HELMHOLTZ-b0.7.0-20m)

1. HELMHOLTZ-b0.7.0-20m: XL-smol World-Model & System-1 Judge Verification (12,003 Held-Out Records / 29,609 Questions)

2. Full 12-Domain XL-smol Evaluation (17,130 Held-Out Test Records / 44,287 Questions)

3. 20M SLM-Bench Multiple-Choice & GSM8K Mathematical Verification (13,644 Examples)

Coming Next: Scaled Rotational JEPA World Models (HELMHOLTZ-100M & HELMHOLTZ-350M)

MatterWaveAI Rotational JEPA Scaling Ladder (10M / 20M / 30M)