AptAI aptai

A centralized adapter registry for foundation models

Discover production-ready fine-tunes, test them instantly in the browser, and deploy them to serverless API endpoints in one click.

Join the Waitlist
AptAI AptAI
Search the registry Ctrl+K
discover / explore adapters
Search the registry
Trending this week
All base models
Filters

Search results

191 adapters match your search
Clear search ✕
Search registry
191 Adapters
adpt-llama-3-8b-instruct
by meta / instruction-tuned
Rank 64
Latency 12ms TTFT
Base Llama-3-8B
Downloads 84.2k
Test in Demo 1-Click Deploy

3 steps. Discover, test, and deploy production-ready fine-tunes in minutes.

Discover specialized adapters
Step 01 / Discover

Discover

Find highly specialized adapters for coding, customer support, data extraction, and more, all built by top creators.

A/B test adapters in side-by-side playground
Step 02 / Test

Test

Stop guessing. Use our side-by-side Playground to instantly A/B test an adapter against its base model before you deploy.

One-click deploy to serverless endpoint
Step 03 / Deploy

Deploy

One click to push any adapter to a production-ready, serverless API endpoint. No GPU provisioning required.

Deploy endpoints at scale with zero cold starts

Router Layer

Dynamic header-based routing matches incoming API requests to adapter registers. Weights swap dynamically inside warm GPU memory pools with sub-millisecond execution overhead, completely eliminating container cold starts.

AptAI AptAI
Search the registry Ctrl+K
infrastructure / deployments

Active Endpoints

3 endpoints routing 14 active adapters
Endpoint Base Model Avg Latency Status
api.aptai.dev/v1/chat 8 adapters
Mistral-7B 1.12ms Active
api.aptai.dev/v1/code 2 adapters
CodeLlama-34B 3.45ms Active
api.aptai.dev/v1/extract 4 adapters
Llama-3-8B 0.84ms Active
api.aptai.dev/v1/embeddings 12 adapters
Nomic-Embed 0.45ms Active
api.aptai.dev/v1/vision 1 adapter
Llava-1.5 — Deploying

Manage Endpoints

Manage and monitor your active deployment routes.
/v1/chat
Llama-3-8B • 8 adapters
142 req/s • 1.12ms
/v1/code
Mistral-7B • 2 adapters
24 req/s • 3.45ms
/v1/vision
Llama-3-Vision • 4 adapters
68 req/s • 1.84ms
/v1/embeddings
Nomic-Embed • 12 adapters
890 req/s • 0.85ms
/v1/extract
Qwen-2-7B • 0 adapters
Offline
/v1/translate
Gemma-2B • 3 adapters
51 req/s • 4.10ms
/v1/chat
Llama-3-8B • 8 adapters
142 req/s • 1.12ms
/v1/code
Mistral-7B • 2 adapters
24 req/s • 3.45ms
/v1/vision
Llama-3-Vision • 4 adapters
68 req/s • 1.84ms
api.aptai.dev/v1/chat
8 active adapters
Throughput 14.2k tok/s Multi-adapter batch
Swap Latency 1.12 ms Zero cold starts
Base Model Llama-3-8B VRAM: 14.8 GB
GPU Utilization 98.4% Dynamic vLLM
Live Dynamic Routing

Hot-swap LoRA adapters on a single base model.

Click an adapter from the registry to route inference to that LoRA in real-time. One base Llama 3.1 8B instance serves 8 distinct personas dynamically with zero container restarts.

Select LoRA Adapter Hot-Swap
aptai playground genz-slang · Llama 3.1 8B

Get paid for your fine-tunes.

Turn your specialized datasets and domain expertise into recurring revenue. AptAI provides the infrastructure to host, protect, and monetize your custom models. Set your own price per token and let thousands of developers route traffic to your endpoint.

70 / 30 Revenue Share

You keep 70% of all inference revenue generated by your adapter. We handle the billing and infrastructure.

Protect Your IP

We serve your adapter on managed endpoints. Your proprietary .safetensors weights are never exposed for public download.

Flexible Pricing

Set your own price per 1M tokens based on the complexity and value of your fine-tune.

Train in minutes. Serve in milliseconds.

Stop wrestling with complex infrastructure. AptAI utilizes state-of-the-art optimizations to make the entire model lifecycle seamless. By leveraging Unsloth-optimized kernels for rapid fine-tuning, and high-density multi-adapter serving (vLLM) on scalable serverless clusters, we deliver the performance of dedicated GPUs at a fraction of the cost.

Infrastructure Metric AptAI Serverless Traditional Hosting
Fine-tuning speed 2x-5x faster (Unsloth-optimized) Hours to days
Adapter density 100+ adapters per base model 1 model per dedicated GPU
GPU Memory Usage Shared base VRAM (Dynamic swapping) Duplicated base footprints
Base hosting costs Pay-per-inference $150+ / mo per dedicated GPU
AptAI AptAI
Search the registry Ctrl+K
studio / fine-tuning / create
Unsaved draft

Fine-Tuning Configuration

Advanced mode: Explicit control over adapter matrices, learning rates, and target layers.

Foundation Model

LoRA Hyperparameters

q_proj k_proj v_proj o_proj gate_proj up_proj down_proj

Dataset

82,419 rows
train_data_v2.jsonl
Max Sequence Length 4096
Pack Sequences

Est. Cost & Time

Hardware 1x H100
Est. Time ~ 34 mins
Total $ 2.45
Run: adpt-llama-8b-vx
Step 120 / 1000
Est. 00:34:00 remaining
Train Loss
0.8423
Eval Loss
0.9104
Learning Rate
1.8e-4
Gradient Norm
2.41
Throughput
14k
GPU VRAM
64.2
Live Adapter Evaluation
Previewing responses with current adapter weights
Updating every 50 steps
"Explain quantum computing in one sentence."
Quantum computing harnesses the laws of quantum mechanics to solve problems too complex for classical computers. It uses qubits which can exist in multiple states simultaneously, allowing for exponential scaling in computational power for specific tasks like prime factorization and complex system simulation.

FAQ

It's a marketplace of production-ready fine-tunes — small, focused model adapters that plug into a shared base model. Instead of hosting a full model per task, you register an adapter once and route requests to it instantly.

Adapters activate inside warm GPU memory pools with sub-millisecond overhead. Because the base model is always resident in VRAM, there is zero container initialization — the first request is just as fast as the hundredth.

Yes. One click pushes any registered adapter to a production-ready serverless API endpoint — OpenAI-compatible, with auto-scaled GPU capacity. No provisioning, no GPU setup.

Every token served through your adapter earns a per-request share, tracked on-chain and paid out automatically. Pricing tiers, royalties, and usage caps are fully yours to configure.

Yes. AptAI acts as an OpenAI-compatible proxy, so OpenHands, AutoGen, CrewAI, and Aider connect with zero code rewrites — just point them at your endpoint.