Skip to content
AI Primer
workflow

Higgsfield runs Fable 5.2 Blender pipeline on five NVIDIA L40S GPUs

Higgsfield shows Fable 5.2 building, rigging and animating Blender characters through its MCP on cloud GPUs. One demo runs the full pipeline on five NVIDIA L40S GPUs.

3 min read
Higgsfield runs Fable 5.2 Blender pipeline on five NVIDIA L40S GPUs
Higgsfield runs Fable 5.2 Blender pipeline on five NVIDIA L40S GPUs

TL;DR

  • Fable 5.2 built, rigged and animated Blender characters through Higgsfield MCP, with Higgsfield's runner demo placing every step on cloud GPUs.
  • Five NVIDIA L40S GPUs powered a character build from base-mesh blockout through 4K textures and final rendering, according to Higgsfield's character demo.
  • A separate local Blender test cost about $3.12 in Astra API calls versus $19 for Fable, but AmirMushich rated Fable's modeling much worse.

Higgsfield already offers Blender presets for exploded-view assembly animations and cartoon shaders. The cloud-computer prompt on its MCP page asks the agent to choose software and hardware, then show the price before starting.

Five L40S GPUs for the Blender pipeline

Higgsfield describes its full Blender character pipeline running on five NVIDIA L40S GPUs. The company says its runner was sculpted, rigged and animated without touching a local GPU.

The fur-collared character build follows these stages:

  1. Base-mesh blockout and voxel sculpting.
  2. Fur collar and 4K textures.
  3. Final rendering.

The posts do not break down which stages used multiple GPUs or give an elapsed time for the complete build.

A 464,581-vertex sculpt

Higgsfield gets pleasingly specific about the geometry in another character breakdown:

  • Complete sculpt: 464,581 vertices.
  • Jacket alone: 56,886 vertices.

The company credits five L40S GPUs with both building and rendering it.

Fable 5.2 through MCP in Claude

For the runner, Higgsfield pairs Fable 5.2 with its MCP connection, while GPU Harness supplies the cloud compute.

A construction-worker demo makes the entry point explicit: Higgsfield MCP inside Claude, with Fable 5.2 credited for the entirely cloud-based build.

A local Blender cost test

In a separate experiment, AmirMushich gave GPT-6.2 Astra and Claude Fable 5.2 the same robotic-hand modeling and animation task in Blender on an RTX 4070 laptop. His reported first-pass API usage was:

| Model | Total tokens | Output tokens | Approximate API cost |
| --- | ---: | ---: | ---: |
| Astra | 3.5 million | 26,000 | $3.12 |
| Fable | 21.2 million | 236,000 | $19 |

His assessment split across three dimensions:

  • Modeling: Fable's hand looked much worse.
  • Movement logic: Fable looked somewhat better.
  • Rendering: He slightly preferred Fable's output.

Fable's output went into procedural components, including circuit boards and actuators, plus checks. AmirMushich said 98% of its tokens were cached context reread at each step.

A second Astra iteration brought its combined usage to 31.3 million tokens. He said that revision improved the render but worsened the animation, and ultimately preferred Astra's first version of the three results.

Render times came with a hardware-load caveat: in his follow-up, Astra's first version took about 21 minutes on an otherwise free laptop, while Fable took 1.5 hours on a busy one.

Cloud access and billing

Higgsfield's GPU Harness page lists /gpu as an entry point and offers a Claude MCP connection. Its changelog describes a configurable cloud computer with preinstalled software, controlled through Claude, ChatGPT, Supercomputer or its desktop from another device.

For Blender workflows, the company describes both finished videos and editable scenes as outputs. Billing is based on time while the computer runs; stopping it saves the workspace for a later session.

The character posts advertise Black Friday discounts, but the fur-collared character post gives no total GPU bill for its build.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 2 threads
Five L40S GPUs for the Blender pipeline2 posts
A local Blender cost test1 post
Share on X