Octane 2027 beta speeds up Gaussian splats by up to 3.8x

On October 31 OTOY released OctaneRender 2027.1 Beta 1 for testing. The main practical news concerns anyone using Gaussian splats in their scenes: according to the company's lab measurements, they render 1.2–2.4 times faster on Windows and Linux and 1.9–3.8 times faster on macOS. In addition, Octane Standalone has gained a built-in MCP server through which AI agents can assemble and edit scenes and launch rendering. The real-time mode has moved to DLSS 4.5.

According to the developers, the 2027 branch is almost feature-complete. New capabilities and stabilization work are still expected before the stable build. OTOY does not advise using the beta on production projects and does not guarantee that scenes saved in it will be compatible with future versions.

Splats got faster, but old scenes may look different

OTOY added support for Gaussian splats with full GPU path tracing back in the stable 2026.1 in November 2025. In 2026.4 it was joined by the SPZv4 format and reduced video memory usage. Alpha 2 brought splat import via glTF 2.0. In Beta 1 the company added a new optimization for rendering them.

OTOY measures the speedup by render time. At the same sample count the noise may be slightly higher than before, but at the same time budget the perceptual error by the FLIP metric dropped in all tests. For a visualization artist who caps rendering by time, that means a cleaner frame in the same number of minutes. Anyone used to setting a fixed sample count will probably have to revisit that threshold.

The speedup has a price. The developers removed the 2026.1 compatibility mode, which reproduced a bug with slightly transparent splats. If a scene from 2026.1 relied on that look of semi-transparent splats, it may appear different in 2027.

Splats have also reached the real-time mode. The same place now lets you enable a new visualization cache.

MCP server: an AI agent drives the scene

Octane Standalone has a Model Context Protocol server built in. Through it a compatible AI agent creates, reviews and edits scenes, sets up materials, lights and cameras and launches rendering on Windows, Linux and macOS. For studios this is a possible basis for automating routine work, such as light placement and typical material tweaks by text command, without separate scripting.

The server has no protection yet, and OTOY says so outright:

"The server uses Streamable HTTP transport, accepts only local connections and has no authentication. Do not expose it to the network."

The agent must run on the same machine as Octane. OTOY does not recommend forwarding the server to a render farm or to a shared studio network.

DLSS 4.5 and denoisers as separate nodes

Real-time neural rendering has been switched to DLSS 4.5. According to OTOY, this delivered more accurate reflections and less ghosting, that is, fewer trails behind moving objects. The update raised the requirements: the minimum NVIDIA driver version is now 580, whereas 2026.4 made do with R555 (R572 for GeForce RTX 50).

Denoisers and upscalers no longer hide in the imager node settings. Each now has its own node:

• denoisers: Octane AI and Open Image Denoise;
• upsamplers: DLSS, FSR, MetalFX, Legacy AI and Linear.

If the required upsampler is not available on the system, for example DLSS on a Mac or MetalFX on Windows, Octane automatically falls back to Linear. In other words, on another platform the scene will be scaled by a simpler method.

Materials, light and USD

• Thin Wall. Rough specular transmission in thin wall mode now accounts for roughness and the index of refraction (IOR). Matte thin surfaces such as glass and films should look more physically accurate.
• Directional light. The angle of a directional source can be set explicitly up to 180 degrees, which allows for softer lighting.
• USD. DistantLight import has been added, so directional sources from USD scenes are now carried over into Octane. A crash on USD export via Geometry Exporter when node names contained slashes or pluses has been fixed.
• MaterialX. On import, Fake Shadows is enabled by default, so direct light passes through transparent materials without manual setup. In MaterialX BSDF layers the separate Opacity input has been removed: layer transparency is now set by the Weight input, as with regular Octane nodes. Old graphs are worth checking.
• Vectron. The Clip and Mirror operators work across several axes at once, so a procedural shape can be mirrored along X and Z with a single operator.

Among the fixes is a bug noticeable on farms: with NRC enabled, network nodes in a multi-GPU configuration dropped to zero GPU utilization.

What the beta lacks

The anime and sketch kernels announced in the Octane 2027 roadmap back in November 2025 did not make it into Beta 1. OTOY confirms they remain in the 2027 release cycle.

The Physical Optics material, first shown in Alpha 2 in July, is also still in progress. The developers are currently moving the optical axis orientation into object space, but do not promise that the remaining issues will be fixed by Beta 2. Sky polarization is not part of the material's scope, and OTOY is not currently working on support for polarized light sources. Physical Optics is slower than regular Specular: it uses a noticeably more complex model with generalized Fresnel calculations, and performance has not been optimized yet.

Requirements and where to get it

You need an NVIDIA GPU of the Turing architecture (compute capability 7.5) or newer and driver version 580 or higher. On Mac, macOS 15 through 27 are supported, whereas 2026.4 ran from macOS 14.5. Studio+ Standalone and Node builds for Windows and Linux have been published, as well as Prime Standalone for macOS. Download links and the full changelog are collected in the OctaneRender 2027.1 Beta 1 thread on the OTOY forum.

What this means in practice

Judging by the contents of the beta, OTOY is betting on two lines: splats as a full-fledged element of a production scene, and render control on the AI side. In 2027 splats already work in both offline and real-time modes. The speedup is most noticeable on Mac, according to OTOY's data.

There are two risks. The first concerns compatibility: the removed 2026.1 mode, the changed logic of the Opacity input in MaterialX and the lack of guarantees for saved scenes make the beta a tool for trying out on copies of projects; it is not meant for live jobs. The second is tied to hardware and environment: the mandatory driver 580 and macOS 15 will probably cut off machines that studios keep on old, proven drivers and systems. The MCP server without authentication is fine for experiments on a workstation, but for integration into a shared studio pipeline it still lacks basic protection.

What would you try first from the Octane 2027 beta?

• Accelerated splats
• The MCP server for AI agents