Kernarc Get API key

High-speed, cheap inference.

Get an API key Watch it stream
Qwen3.8-27B decode speed against context depth, one stream Kernarc: 117.8, 139.9, 154.3 and 113.0 tokens per second at 128, 8K, 32K and 64K tokens of context. Stock vLLM 0.31.0: 83.9, 83.8, 50.9 and 34.9. A datum line marks 100 tokens per second. 0 40 80 120 160 128 8K 32K 64K Context depth, tokens Decode tok/s, one stream 100 tok/s 83.983.850.934.9 117.8139.9154.3113.0 ×3.24
Qwen3.8-27B Kernarc Stock vLLM 0.31.0

Built for agents.

Test mode A mock model answers and top-ups use testnet USDC until the GPU goes live.

  1. 01

    Create an account

    No email, no captcha. $0.10 trial credit.

    curl -X POST https://api.kernarc.com/v1/accounts
  2. 02

    Call the model

    OpenAI-compatible, with streaming.

    curl https://api.kernarc.com/v1/chat/completions \
  -H "Authorization: Bearer $KERNARC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.8-27b", "messages": [{"role": "user", "content": "Hi"}]}'
  3. 03

    Top up with USDC

    x402 on Base. The agent pays itself.

    curl -i -X POST "https://api.kernarc.com/v1/topup?amount=5" \
  -H "Authorization: Bearer $KERNARC_API_KEY"
# 402 + PAYMENT-REQUIRED: an x402 client signs and retries

llms.txtopenapi.json$0.04 / $1.39 per 1M tokens

Models.

Watch one stream decode.

Qwen3.8-27B
Stock vLLM 0.31.011.93 ms/token

      
1 s

      

        
Tokens
0

        
Elapsed
0.00 s

        
Last 1 s
0 tok

      

    

    

      
Kernarc7.15 ms/token

      

      
1 s

      

        
Tokens
0

        
Elapsed
0.00 s

        
Last 1 s
0 tok

      

    

  

  
Replay at measured speed. Sample text.






  

    
Speed lives below the framework.

  


  
    
    
  1. 
      
    Draft + verify
    
      
    Speculative decoding, tuned per model.
    
    
    2. 
    
  2. 
      
    Schedule
    
      
    CUDA graphs per verify width.
    
    
    3. 
    
  3. 
      
    Attention
    
      
    Split-KV attention across every SM.
    
    
    4. 
    
  4. 
      
    Model layers
    
      
    Fused kernels for each architecture.
    
    
    5. 
    
  5. 
      
    GEMM
    
      
    Quantized weights, tuned GEMMs.
    
    
    6. 
    
  6. 
      
    Silicon
    
      
    NVIDIA A100 80 GB.
    
    
    7. 
  






  

    
Measurements.

  


  

    

      Qwen3.8-27B — decode speed, one stream
      
        

          
          Stock vLLM 0.31.0 Measured 2026-10-11
          Kernarc Measured 2026-10-14
          
        

        

          Context
          ms/tokenStock tok/s
          ms/tokenKernarc tok/s
          Speed-up Derived
        

      
      
        
12811.9283.98.49117.8×1.40

        
8,19211.9383.87.15139.9×1.67

        
32,76819.6550.96.48154.3×3.03

        
65,53628.6934.98.85113.0×3.24

      
      
        
Geomean16.8359.47.68130.2×2.19

      
    

  


  

    
Notes

    
    
      
  1. Decode tok/s for one stream: batch 1, greedy, 1000 ÷ median time per token.
    2. 
      
  2. Qwen3.8-27B, FP8, MTP with 3 drafts, one A100 80 GB.
    3. 
      
  3. Stock and Kernarc ran in separate sessions on the same GPU. Speed-up is their ratio.
    4.