$$$$$$@@@@@@@@@@                                                                          
                                                                  ###########$$$$$$$$$$$$$                                                                     
                                                                **************#######$$$$$$$$                                                                  
                                                               !!=!!!!!!!!*!*!******########$$#                                                                
                                                              ;;============!!!!!*!*******#######                                                              
                                                              ;:::::::::;;;;=====!!!!!!!*!*****#*#*                                                            
                                                              ~~~--~~~~~~~~~::;;;;;===!!!!!!*!******                                                           
                                                              -,,,.....,,,---~~~::;:;;=====!!!!!!!!!!                                                          
                                                              ,..............,,---~~:::;;;=;====!!!=!!                                                         
                                                               ..................,---~~~:::;;;;========                                                        
                                                                ....................,,,-~~~~::;;;;;;;=;                                                        
                                                                  ........,~=*#$.......,,--~~~~:::;;;;;                                                        
                                                                   ......-~;=*$$$$........,,--~~~::::::                                                        
                                                                      ..,-~;!!****=:........,,---~~~~~~                                                        
                                                                        .,-:;=!!=;::,.........,,,-----                                                         
                                                                           ..-::;;~~,.............,,,                                                          
                                                                               ..,,,...............                                                            
                                                                                    ...........

$ memora recall "apple silic

Hoshi Labs · Est. 2024 · Tokyo

AI Infrastructure
for Apple Silicon.

Mechanistic interoperability. Frontier research.
We build on the compute nobody else uses.

Explore Products GitHub
● ANE_UTIL: 89%PROOF_HASH: 0x3f8a..c7d2✦ MEMORA: 1,247 memories indexedMTP_ACC: 86% · SPEEDUP: 2.89×EPOCH: 1030 · CHAIN: sui_testnet● NODES: active · TOK/S: 234CHIMERA: mtp.0 weights restoredINFERENCE_PROOF submitted on-chain✦ ANE: 178 calls/sec/W · GPU: 12AGENT[reasoning]: Qwen3-0.6B · WARM● ANE_UTIL: 89%PROOF_HASH: 0x3f8a..c7d2✦ MEMORA: 1,247 memories indexedMTP_ACC: 86% · SPEEDUP: 2.89×EPOCH: 1030 · CHAIN: sui_testnet● NODES: active · TOK/S: 234CHIMERA: mtp.0 weights restoredINFERENCE_PROOF submitted on-chain✦ ANE: 178 calls/sec/W · GPU: 12AGENT[reasoning]: Qwen3-0.6B · WARM
Scroll
Your Mac has a Neural Engine. Apple built it. Nobody uses it. We changed that.
0×
More power efficient than GPU ANE vs Metal for MTP inference on Apple Silicon. From our Chimera research.

What we build

Four products.
One substrate.

Hoshi Node

Your Mac
earns crypto.

Real MLX inference on Apple Silicon. Every job produces a cryptographic proof, anchored on Sui. Early node operators join now.

On-chain InferenceProof via Move contract
zkLogin — Sign in with Apple, no seed phrase
4 specialized ANE agents per node
Early Access 

    


    
    

      

        

        
Memora

        
Memory that
never forgets.

        
Local-first memory layer for AI agents. ANE-accelerated embeddings. Zero API keys, zero cloud. Drop-in for any agent framework.

        

          
0.97ms embed latency on ANE

          
100% recall at 1M memories

          
Bio-inspired short → long → core tiers

        

        
          View on GitHub
          
        
      

      

    


    
    

      

        

        
Chimera

        
2.89× faster
inference.

        
First implementation of MTP inference on MLX. We re-enabled the speculative decoding weights every framework strips. Published results.

        

          
76–86% MTP accuracy on Qwen3.5

          
15× more efficient than GPU on ANE

          
234 tok/s on M3 Ultra

        

        
          Read the findings
          
        
      

      

    


    
    

      

        

        
Hawkeye

        
Search that
sees everything.

        
Parallel multi-LLM research engine. Dispatches queries to 7 providers simultaneously — Perplexity, Tavily, DeepSeek, Gemini, and more. Synthesizes consensus in seconds.

        

          
7 LLM providers in parallel

          
~$0.05 per research query

          
Rust CLI — instant consensus synthesis

        

        
          View on GitHub
          
        
      

      

    


  







  

    

      
Open Research

      
We publish
what we find.

      

        The ANE is the most underutilized compute substrate in the world. We benchmark it, break it, and publish the results. No paywalls, no preprints — just findings on GitHub.
      

      

        
          research@hoshilabs.co
          
        
      

    


    

      

        

          F·01
          

            
MTP weights survive 4-bit quantization

            
Qwen3.5 MTP heads achieve 76–86% accuracy after 4-bit quantization with fp16 sidecar. mtp.* weights stripped by MLX were never the bottleneck.

          

        

        

          F·02
          

            
RoPE encoding is load-bearing for MTP

            
Without positional encoding on the MTP head, accuracy collapses to 0%. Every framework that strips these weights also strips RoPE. Both matter.

          

        

        

          F·03
          

            
ANE runs 178 calls/sec/W vs GPU's 12

            
For MTP heads on Apple Silicon: ANE at ~3W delivers 15× the power efficiency of Metal GPU. The chip Apple built for ML is the right chip for ML.

          

        

        

          F·04
          

            
Fire-and-forget ANE swarms are 14% faster

            
127 specialist models running simultaneously with no coordination overhead outperform sequential dispatch by 14%. Discovered during Hoshi Engine benchmarking.

          

        

      


      

        
Compute Efficiency — ANE vs GPU vs CPU

        

          

            

              ANE (Neural Engine)
              178 calls/sec/W
            

            

              

            

          

          

            

              GPU (Metal)
              12 calls/sec/W
            

            

              

            

          

          

            

              CPU
              6 calls/sec/W
            

            

              

            

          

        

        
M3 Ultra · MTP head inference · Chimera research, 2026