From Hugging Face to the Corsair in Hours

Bringing a new model to a new accelerator typically takes weeks to months of engineering effort.

On Tuesday, October 6, at 5 p.m. PT, we put Ignition, our hardware enablement agent, to work on the d-Matrix Corsair.

By Sunday, October 11, at 5 p.m. PT, five days later, Ignition had built and extended a custom compiler, surrounding tooling and verification pipeline to enable 39 models spanning language, vision, speech and image generation.

With that infrastructure in place, bringing a new model to the Corsair now takes hours.

Enabled by Ignition

15 language-model families. Vision. Speech. Image generation.

Across tested Corsair configurations, Ignition achieved performance comparable to our hand-tuned reference implementation, previously accepted by the d-Matrix team.

Text

33 Text Models

• Language

  Gemma 4 12B

• Language

  Qwen3-0.6B

• Language

  Sarashina2.2-3B

• Language · MoE

  Granite-3.1-1B-A400M

• Language

  Gemma-3-1B

• Language

  Qwen3.5-9B

• Language

  Sarashina2.2-1B

• Language · MoE

  OLMoE-1B-7B

• Language · MoE

  Qwen1.5-MoE-A2.7B

• Language

  Llama-3.2-1B

• Language

  EXAONE-4.0-1.2B

• Language

  OLMo-2-1B

• Language

  Gemma-2-2B

• Language

  Gemma-4-E2B

• Language

  Qwen3.5-0.8B

• Language

  Llama-3.2-3B

• Language

  Llama-3.1-8B

• Language

  TinyLlama-1.1B

• Language

  Mistral-7B-v0.3

• Language

  Ministral-8B

• Language

  Qwen2.5-7B

• Language

  Qwen3-4B

• Language

  Phi-3.5-mini

• Language

  Phi-4-mini

• Language

  Gemma-3-4B

• Language

  SmolLM3-3B

• Language

  StableLM-2-1.6B

• Language

  Falcon3-1B

• Language

  Granite-3.3-2B

• Language

  GLM-4-9B

• Language

  OLMo-3-7B

• Language

  Hunyuan-0.5B

• Language

  ERNIE-4.5-0.3B

Beyond Text

6 Models Across Modalities

• Vision + language

  LLaVA-1.5-7B

• Speech recognition

  Whisper large-v3

• Speech synthesis

  Kokoro-82M

• Image classification

  ResNet-50

• Image generation

  Stable Diffusion v1.5

• Vision + language

  Qwen2-VL-7B

IGNITION ON CORSAIR

Build the Capability. Enable the Next Model.

THE INITIAL ENABLEMENT

5 Days

Tue, Oct 6, 5 p.m. → Sun, Oct 11, 5 p.m. PT

Ignition built and extended a custom compiler, tooling and verification pipeline, achieving the coverage above.

A NEW MODEL, NOW

Hours

Ignition uses and extends that infrastructure to bring up the next model.

The Agent Builds the Capability

Ignition does the engineering. The compiler is one of its tools.

Ignition builds the infrastructure to map models onto the Corsair, run them on cards and check their outputs.

When a new model or model family needs additional capabilities, Ignition extends the compiler and tooling automatically, verifies the changes on hardware, and checks that existing models still work. Each improvement carries forward.

Ignition

Hardware Enablement Agent

Builds, extends and operates the toolchain below.

1. 1 / START

   Hugging Face

   Checkpoint + target configuration
2. 2 / COMPILE

   Custom Compiler

   Place work. Split layers. Move data.
3. 3 / RUN

   Corsair Cards

   Execute the compiled program on hardware.
4. 4 / VERIFY

   Check the Outputs

   Bit-exact execution + original-model comparison

Give Ignition a Hugging Face checkpoint and a target: batch size, context length and card count. It uses that infrastructure to bring up the model in hours.

One Model, Many Deployments

Ignition adapts models to different deployment goals. Depending on the model, tested configurations include:

• Fewer cards: compressed weights enable deployments on as few as 2 cards.
• More concurrency: up to 16 concurrent sequences on 4 cards.
• Longer context: a single sequence at 64K context on 8 or 16 cards.
• Flexible batch sizes: batch 1 and batch 4 with compressed weights on 8 cards.

Compressed weights trade some accuracy for capacity: larger models on fewer cards, or longer contexts on the same hardware.

Accuracy, Measured

Hardware execution matches the compiled program's software model bit for bit.

Quantization and numerical data types introduce differences from the original Hugging Face model. Ignition measures those differences separately.

• Text: 92–97% exact token agreement in standard formats; 85% for Gemma 4 12B with compressed weights on 8 cards.
• Vision-language: 89–97% exact token agreement.
• Whisper: word-for-word transcript agreement on our test clips.