NVIDIA published a developer post on October 27, 2026 describing what game developers can use on RTX Spark, its Windows PC platform for slim laptops and compact desktops. The platform pairs a Blackwell RTX GPU with a Grace CPU and up to 128 GB of unified memory, per NVIDIA's technical blog. Every performance number below is NVIDIA's own; we have not benchmarked any of it.
If you want the platform background first, see our posts on RTX Spark launch software and using RTX Spark for local coding agents.
Quick reference
| Item | What NVIDIA says |
|---|---|
| DLSS 5 on RTX Spark | 3D-Guided Neural Rendering supported, starting with NBA 2K27 |
| October DLSS 5 model | 2x faster than September, half the GPU overhead, 20% less VRAM at 4K |
| Nemotron 3.5 Bolt E4B 4GB | About 40% higher average quality than prior model, up to about 80% fewer tokens vs peers |
| RTX Kit | Eight SDKs published with RTX Spark support |
| Tools | Nsight Graphics 2026.3, Nsight Systems 2026.5, Aftermath SDK 2026.3 for Windows on Arm |
| Engine | NVIDIA RTX Branch of Unreal Engine 5.8; DLSS 4.5 plugin 8.9.0 with Windows on Arm support |
| Motion | MotionBricks adds SOMA human-motion models; research code only |
DLSS 5 with 3D-Guided Neural Rendering
NVIDIA says DLSS 5 adds a final stage to the graphics pipeline (see NVIDIA's earlier game-developer update) that enhances lighting and material detail. The engine-rendered frame, including artist-designed geometry, textures and lighting buffers, stays the foundation, and developer controls decide what can change. Avalanche Software used those SDK controls to keep the artistic intent of Hogwarts Legacy, which NVIDIA shows in a side-by-side.
The efficiency claim is the news: an October model update that NVIDIA says runs 2x faster than September's, halves GPU overhead at comparable fidelity and trims 4K VRAM use by 20%. It will ship by default in future DLSS 5 games. DLSS 5 was controversial when shown at SIGGRAPH, which we covered in the AI slop debate post; the developer-control framing here is NVIDIA's answer to that. For other uses of the technique see neural rendering use cases beyond DLSS 5.
NVIDIA's earlier developer update describes the technique in more detail: it works one frame in, one frame out using motion vectors, runs locally on GeForce RTX 50 Series GPUs at up to 4K, and gives developers controls for model choice, effect intensity and AI-based or engine-level masking. NBA 2K27 was the first game named as supporting it. For a studio, the practical question is how much of the frame the model is allowed to touch, and those masking controls are where that decision gets made.
The post also lists the rest of the DLSS suite on RTX Spark: Super Resolution, Ray Reconstruction and Dynamic Multi Frame Generation, which can generate up to five extra frames per rendered frame. The DLSS 4.5 Unreal Engine plugin 8.9.0 adds official Windows on Arm support for UE 5.8, giving a supported path to native Arm64 builds.
Nemotron 3.5 Bolt E4B for on-device characters
NVIDIA ACE is its suite for in-game AI agents. The post cites two showcases: PUBG's AI teammates in public beta on a first-generation Nemotron small model, and Aniimo's voice-responsive pets on a second-generation one, due in early 2027.
The new piece is Nemotron 3.5 Bolt E4B 4GB, built for character interaction with multi-turn tool use, a conversational mode for short speech-ready replies, configurable reasoning, 64k context and 35 languages. It is meant to leave memory headroom for graphics and local speech models. We covered the model itself in Nemotron 3.5 Bolt E4B: Nvidia's 4GB open edge model.
Reading NVIDIA's chart: it plots an average Q4_K_M score across MultiChallenge, RoleMRC, BFCL-v3, IFEval, SysBench and RelyToolBench, measured on an RTX 5090. Nemotron 3.5 Bolt E4B scores 70.2 at roughly 300 ms time to first token. Gemma 4 26B-A4B scores 72.8 but is slower and uses much more VRAM; Gemma 4 E4B scores 63.2 and Qwen 3.5 9B 64.1. The chart's cloud reference points are Gemini 3.8 Flash at 77.1 (2,449.5 ms) and Claude Sonnet 5.5 at 76.2 (3,002.2 ms). So the honest summary is that a 4 GB local model gets within a few points of a much larger local one at lower latency, while still trailing the cloud models on quality. These are vendor-selected benchmarks on vendor hardware.
How to try it
Developers get the model through NVIDIA's In-Game Inferencing (IGI) SDK. For a quick hands-on test, NVIDIA points to SkyrimNET, a mod platform by MinLL that turns every Skyrim NPC into an unscripted character with memory, voice and autonomy, supporting cloud and local inference. An ARM-supported Mod Organizer loader is linked for RTX Spark. Expect a mod, not a polished product, and expect that latency and VRAM budgets inside a real game differ from a benchmark chart.
RTX Kit: eight neural rendering SDKs
NVIDIA lists eight RTX Kit technologies with RTX Spark support:
- RTX Neural Texture Compression: AI-shrunk texture storage at fidelity comparable to block compression.
- RTX Texture Filtering: stochastic sampling after shading to reduce artifacts.
- RTX Neural Shaders: train and run small networks inside shaders.
- RTX Mega Geometry: faster BVH construction for cluster-based geometry.
- NVIDIA Real-Time Denoisers (NRD): cleaner images from few rays per pixel.
- RTX Character Rendering: strand-based hair and skin with path tracing.
- RTX Global Illumination (RTXGI): efficient indirect lighting.
- Opacity Micro-Maps (OMM): faster ray tracing of high-detail geometry.
Developers can start with the RTX Neural Texture Compression repository on GitHub, per NVIDIA's next steps.
Tools: Arm64 from day one
Windows on Arm is the practical catch for RTX Spark. NVIDIA says its Nsight tools handle native Arm64, emulated x64 and Arm64EC apps from day one: Nsight Graphics 2026.3 adds a Windows on Arm installer, Nsight Systems 2026.5 has one installer for all Windows versions, and the Nsight Aftermath SDK 2026.3 has an Arm version for crash dumps (emulated x64 apps keep using the x64 SDK). The NVIDIA RTX Branch of Unreal Engine 5.8 focuses on refactoring, stabilization and compatibility.
MotionBricks
MotionBricks is open-source research for real-time character motion: you give target poses and it generates the motion between them at runtime, with control over speed, direction and style, plus interactions like sitting and picking up objects. The update adds SOMA human-motion models alongside the existing G1 robot model, downloadable checkpoints, ONNX exports, export tools and an interactive reference built with Newton. The Unreal Engine integration from the original demo is not included. See our earlier MotionBricks coverage.
Why local AI characters are hard in a shipping game
A game already spends most of a GPU's memory and time on rendering. Adding a language model for NPC dialogue competes for exactly the same VRAM and compute, so the design goal NVIDIA states, maximum intelligence and speed per gigabyte, is the right one. A 4 GB model leaves room for textures; a 26B model does not, which is why the comparison with Gemma 4 26B in NVIDIA's chart is framed around VRAM as well as score.
Three practical problems remain that the post does not claim to solve. First, latency under load: a time to first token measured on an otherwise idle RTX 5090 will not hold while the GPU is also rendering at 4K with frame generation. Second, consistency: an unscripted NPC can contradict quest logic or say something out of tone, so developers need guardrails and grounding to game state, which NVIDIA lists as a model capability but which still has to be integrated per game. Third, shipping and licensing: every title that bundles a model has to account for download size, memory on lower-end machines, and fallback behavior when the model is not available.
The PUBG and Aniimo examples suggest NVIDIA's own expectation: narrow, well-bounded companions (a teammate, a pet) before open-ended conversation with every villager. SkyrimNET is the opposite end of that spectrum, a community project that tries open-ended conversation everywhere, and is a useful place to see both the promise and the rough edges.
Why this matters beyond games
RTX Spark is positioned as a platform for AI development, creation and gaming alike. The game-developer post is partly a signal that NVIDIA wants Arm-based Windows PCs to be taken seriously as a target, with the same CUDA stack and RTX features developers already know. If the tooling holds up, the same machine that runs a local coding agent can profile a game frame, which is the case our RTX Spark coding-agent post makes from the other direction.
What this means for developers
- Check the Arm story first. If your engine and middleware do not build for Windows on Arm, the AI features matter less.
- Budget memory. The pitch for Bolt E4B is intelligence per GB; verify it against your actual VRAM headroom.
- Treat percentages as vendor claims. The 2x and 20% figures describe NVIDIA's own model versions; test on your content.
- Decide your DLSS 5 stance early. Controls exist to constrain what the network changes, which matters for art direction.
Related reading
- RTX Spark launch software: Nemotron 3.5 Bolt, PAIR 1.0 and Llama Windows
- Nemotron 3.5 Bolt E4B
- RTX Spark for local coding agents
- DLSS 5 and the AI slop debate
- Neural rendering use cases beyond DLSS 5
- NVIDIA MotionBricks
Details are accurate as of October 27, 2026; all performance figures are NVIDIA-reported.