SIMFORGE-D1.5

SIMFORGE-D1.5 is a compact vision-language-action driving model distilled from NVIDIA's Alpamayo 1.5 10B. Its primary goal is edge deployment: preserving the parent's driving behavior while reducing the model enough to make Alpamayo-class inference practical on constrained accelerators such as NVIDIA Jetson AGX Orin-class systems and discrete 16 GB GPUs.

Introduction

We introduce SIMFORGE-D1.5, a 7.21B-parameter distilled checkpoint derived from the 11.08B-parameter Alpamayo 1.5 model. SIMFORGE-D1.5 is approximately 35% smaller than its parent and reached a 14.42 GiB peak model reservation in our deployment evaluation, placing the model itself within a 16 GiB GPU memory envelope.

The checkpoint was produced through staged model reduction and distillation specifically to move an Alpamayo-class driving model from datacenter-only hardware toward edge-class deployment. Candidate components were removed incrementally, the smaller model was trained to reproduce the parent's driving and reasoning behavior, and each stage was evaluated before selecting this checkpoint as the best size-quality tradeoff in the study.

The 14.42 GiB result was measured on an NVIDIA A100. It establishes model-memory fit within 16 GiB, not complete platform compatibility: Jetson AGX Orin uses shared system memory, and its runtime kernels, compiler support, operating-system allocation, and end-to-end latency still require device-specific validation.

In our 94-scene evaluation, SIMFORGE-D1.5, Alpamayo 1.5, and FlashDrive performed similarly overall. Because this was a single rollout per scene and results varied across scenes, we do not consider the observed score differences strong enough to establish a clear performance difference between the models.

Results

We evaluated SIMFORGE-D1.5 against stock Alpamayo 1.5 and Alpamayo 1.5 with compiled FlashDrive on 94 NuRec scenes. Every arm ran the same scenes once with AlpaSim 0.89.0 and the e2e_challenge=dev protocol.

Metric	Alpamayo 1.5 10B	SIMFORGE-D1.5	Alpamayo 1.5 + FlashDrive
Mean AlpaSim scene_score ↑	0.6500	0.6229	0.7221
Median scene_score ↑	0.8764	0.7846	1.0000
Pass rate ↑	75.53%	74.47%	77.66%
At-fault collisions ↓	8	10	6
Off-road failures ↓	15	14	15
Total driving failures ↓	23	24	21
Peak model reservation	21.99 GiB	14.42 GiB	20.81 GiB

SIMFORGE-D1.5 retained 95.8% of the parent's observed mean scene score while using the smaller checkpoint. Overall, the three systems delivered similar closed-loop driving performance in this evaluation, and the observed differences are not strong enough to support a definitive ranking. FlashDrive is a compiled asynchronous deployment of Alpamayo 1.5, not a separate distilled checkpoint.

Controlled Model Latency

Controlled model latency is the time the model itself needs to produce a driving plan. We gave every model the same four-camera input on the same NVIDIA A100 40GB and timed from JPEG decoding through trajectory output. Lower is faster. Three warmups were excluded, followed by three measured runs.

Metric	Alpamayo 1.5 10B	SIMFORGE-D1.5	Alpamayo 1.5 + FlashDrive
Synchronous latency samples	1.687 / 1.693 / 1.699 s	1.468 / 1.556 / 1.569 s	1.254 / 1.238 / 1.262 s
Median synchronous latency ↓	1.693 s	1.556 s	1.254 s
Latency reduction vs. parent	—	8.1%	25.9%

SIMFORGE-D1.5 reduced median synchronous latency by 8.1% relative to stock Alpamayo 1.5. FlashDrive was measured in its deployed warm steady-state streaming mode; compilation and initial prefill were excluded.

Deployment RPC Latency

Deployment RPC latency is how long AlpaSim waits for the driver to answer a Drive request. It measures the serving system, not just model computation.

Metric	Alpamayo 1.5 10B	SIMFORGE-D1.5	Alpamayo 1.5 + FlashDrive
Drive calls	17,099	18,697	18,697
Pooled mean RPC latency ↓	1747.75 ms	2.99 ms	3.02 ms
Median scene-mean RPC latency ↓	1722.56 ms	2.79 ms	2.81 ms
P95 scene-mean RPC latency ↓	1960.71 ms	3.83 ms	4.03 ms

Stock Alpamayo runs the model before replying, so its RPC latency is close to its model latency. SIMFORGE-D1.5 and FlashDrive run the model in the background and immediately return the latest cached plan. Their roughly 3 ms RPC latency means the simulator receives a response quickly; it does not mean the model finishes inference in 3 ms.

Evaluation Setup

• AlpaSim 0.89.0 with e2e_challenge=dev
• One rollout per scene
• 200 simulation steps
• 1.7-second forced-ground-truth interval
• 100 ms control timestep
• 40 m route-start offset
• Four 1080p camera observations
• One NVIDIA A100 40GB per worker
• Evaluation manifest SHA256: 388d30b5a2d7106df1a5a98b5b2ad62298bd63a22623906b117727c9a63ed6c3

The benchmark uses one rollout per scene. Paired intervals quantify variation across scenes, not run-to-run stochasticity. The challenge uses a private scene distribution, so NuRec96 is an internal comparative benchmark rather than a challenge-score estimate. Because the base and optimized arms use different serving architectures, these results support deployment-level conclusions rather than weights-only superiority claims.

Machine-readable results, per-arm scorer output, and the exact manifest are available under benchmark/nurec96/.

Intended Use

SIMFORGE-D1.5 is released for autonomous-driving research, model-compression experiments, simulation evaluation, and deployment engineering. It is not a production driving system and must not be used to control a real vehicle or make safety-critical decisions.

License and Attribution

SIMFORGE-D1.5 is derived from NVIDIA Alpamayo 1.5. The model weights are provided under the included OpenMDW 1.1 license. Users are responsible for reviewing the upstream model terms and applicable dataset licenses before use. NVIDIA has not endorsed this release.