Sustainable AI Inference

Cobble

Cobble is NOT a datacenter

AI inference that doesn't cost the earth. Built from reclaimed GPUs and server hardware — extending useful life instead of manufacturing new silicon. Deployed in regions with access to renewable energy.

Curated catalog · Quantized & benchmarked
Generative AI

Qwen3.8 27B

FLAGSHIP
Generative AI

DeepSeek V4 Flash

FEATURED
Generative AI

Qwen3.8 Flash Next

Generative AI

Qwen3.6 35B A3B

Generative AI

Qwen3.5 9B

Generative AI

Gemma4 31B

Generative AI

Gemma4 26B A4B

Generative AI

Gemma4 12B

Generative AI

Gemma4 E4B

Generative AI

Gemma4 E2B

Generative AI

Ornith 1.5 35B A3B

Generative AI

Ornith 1.5 9B

Generative AI

Laguna S 2.1

Generative AI

MiMo V2.6 Distill 9B

Generative AI

Ling 3.0 Tiny

Generative AI

LFM2.5 8B A1B

Generative AI

Granite 4.1 8B

Generative AI

Granite 4.0 H Tiny

Generative AI

Mistral Nemo 12B

Generative AI

GPT-OSS 20B

Generative AI

Muse Glimmer 30B

Generative AI

Hermes Compressor

OCR

GLM-OCR

OCR

DeepSeek OCR2

Embeddings

EmbeddingGemma 300M

FEATURED
Generative AI

Qwen3.8 27B

FLAGSHIP
Generative AI

DeepSeek V4 Flash

FEATURED
Generative AI

Qwen3.8 Flash Next

Generative AI

Qwen3.6 35B A3B

Generative AI

Qwen3.5 9B

Generative AI

Gemma4 31B

Generative AI

Gemma4 26B A4B

Generative AI

Gemma4 12B

Generative AI

Gemma4 E4B

Generative AI

Gemma4 E2B

Generative AI

Ornith 1.5 35B A3B

Generative AI

Ornith 1.5 9B

Generative AI

Laguna S 2.1

Generative AI

MiMo V2.6 Distill 9B

Generative AI

Ling 3.0 Tiny

Generative AI

LFM2.5 8B A1B

Generative AI

Granite 4.1 8B

Generative AI

Granite 4.0 H Tiny

Generative AI

Mistral Nemo 12B

Generative AI

GPT-OSS 20B

Generative AI

Muse Glimmer 30B

Generative AI

Hermes Compressor

OCR

GLM-OCR

OCR

DeepSeek OCR2

Embeddings

EmbeddingGemma 300M

FEATURED
The Cobble difference

Not your typical inference provider

Traditional AI providers burn megawatts in massive datacenters. We built something different.

vs
Traditional
Infrastructure
Massive datacenters
Distributed edge nodes
Power Source
Grid-dependent megawatts
Renewable-ready regions
Cooling
Evaporative water cooling
No evaporative cooling
Hardware
Proprietary enterprise GPUs
Reclaimed GPUs & servers
Carbon Footprint
High manufacturing churn
No new-silicon manufacturing
Model Focus
Full precision only
Per-model quantization
Receipts, not promises

Built to do better

Every component was sourced, recycled, and repurposed.

0
Water Usage
0%
Green Energy
0%
Recycled Hardware
-0x
Carbon Footprint
Open numbers · Open weights · Open methodology

Quantized. Benchmarked. Real.

We publish what others hide. Every model is tested, quantized, and documented.

Qwen3.8 27B

FP8
Context
128K tokens
Throughput
55 tokens/sec

DeepSeek V4 Flash

FP8
Context
256K tokens
Throughput
See catalog

EmbeddingGemma 300M

FP8
Context
2K tokens
Throughput
See catalog
Three steps to inference

How it works

01

Choose a model

Pick from our curated selection of quantized models optimized for speed and quality on recycled hardware.

02

Send a request

Use our OpenAI-compatible API. Drop-in replacement for your existing inference pipeline.

03

Get results

OpenAI-compatible responses from distributed edge nodes running vLLM.

Ready to route smarter?

Join the inference network that gives back to the planet.