Discussions

• 1 You must be logged in to vote

  💬

  8xRTX 5090, very cheap and very fast

  horsten started Mar 22, 2026 in General

  

  

  3
• 1 You must be logged in to vote

  💡

  try xSILU (Expanded Silu) and xSwiglu (Expanded swiglu)

  Tqela started Mar 23, 2026 in Ideas

  

  

  2
• 1 You must be logged in to vote

  🙏

  7B run with current version - any gotchas before I commit the budget?

  Muesli1 asked Mar 10, 2026 in Q&A · Unanswered

  

  

  1
• 229 You must be logged in to vote

  📣

  Introducing nanochat: The best ChatGPT that $100 can buy.

  update

  karpathy announced Oct 13, 2025 in Announcements

  

  

  

  

  97
• 1 You must be logged in to vote

  💬

  Successful run on vast.ai

  pradiptamitra started Mar 21, 2026 in General

  

  

  0
• 1 You must be logged in to vote

  🙌

  161 experiments on RTX 3060: optimal architecture shifts under step starvation

  ATSpuler started Mar 17, 2026 in Show and tell

  

  

  2
• 1 You must be logged in to vote

  💬

  Nanochat tool calling

  mrsipan started Mar 17, 2026 in General

  

  

  0
• 2 You must be logged in to vote

  💬

  Pretty impessive for a d12 model trained in 144m + 63m SFT on a single RTX 5090

  horsten started Mar 14, 2026 in General

  

  

  0
• 126 You must be logged in to vote

  📣

  Beating GPT-2 for <<$100: the nanochat journey

  update

  karpathy announced Jan 31, 2026 in Announcements

  

  

  

  

  41
• 1 You must be logged in to vote

  🙌

  Game theoretical pressure and metaphysical grounding eliminate hallucinations and drift from the response; zer failure rate

  ediestel started Feb 23, 2026 in Show and tell

  

  

  1
• 1 You must be logged in to vote

  🙏

  Lambda out of capacity

  pradiptamitra asked Mar 10, 2026 in Q&A · Unanswered

  

  

  

  

  14
• 1 You must be logged in to vote

  🙌

  DCLM-baseline dataset for nanochat (drop-in replacement)

  ddudek started Jan 28, 2026 in Show and tell

  

  

  

  

  2
• 1 You must be logged in to vote

  💡

  Is it time for nanoagent?

  MichelNivard started Mar 8, 2026 in Ideas

  

  

  0
• 1 You must be logged in to vote

  🙏

  scripts.chat_cli only producing <|bos|>

  svlandeg asked Mar 4, 2026 in Q&A · Answered

  

  

  2
• 1 You must be logged in to vote

  🙌

  nanoPLM (Protein language model) and flash-mHC (optimized Triton kernels for deepseek mHC)

  alint77 started Mar 5, 2026 in Show and tell

  

  

  0
• 1 You must be logged in to vote

  🙏

  No masking for response loss in renewed SFT training?

  RmZeta2718 asked Feb 27, 2026 in Q&A · Answered

  

  

  3
• 1 You must be logged in to vote

  🙏

  Why mid-train is removed from repo?

  leolbyy asked Feb 4, 2026 in Q&A · Answered

  

  

  

  

  6
• 2 You must be logged in to vote

  🙌

  NanoKnow: A benchmark to evaluate what your nanochat checkpoint actually knows

  lingwei-gu started Feb 26, 2026 in Show and tell

  

  

  1
• 2 You must be logged in to vote

  💬

  Summary module for NanoChat neural networks.

  JohnDrakopoulos started Feb 24, 2026 in General

  

  

  0
• 1 You must be logged in to vote

  💡

  Universal Intelligence Transfer: Bridging Data-Rich and Low-Resource Domains

  onchiptech started Feb 22, 2026 in Ideas

  

  

  0
• 1 You must be logged in to vote

  💬

  chat_rl not learning <|assistant_end|>?

  lichengunc started Feb 22, 2026 in General

  

  

  1
• 2 You must be logged in to vote

  🙌

  nanollama — the 2026 "learn-from-scratch" baseline (born from nanochat)

  ariannamethod started Feb 22, 2026 in Show and tell

  

  

  0
• 1 You must be logged in to vote

  💬

  Why not use ScatterMoE as the fused MoE kernel?

  serdardoesml started Feb 20, 2026 in General

  

  

  1
• 2 You must be logged in to vote

  💡

  Community engagement / Random questions

  bensonlee5 started Feb 19, 2026 in Ideas

  

  

  0
• 1 You must be logged in to vote

  💡

  Why not just surpass 5.2 directly? Let's add the LiE enhancement plugin.

  ZengHaohei started Feb 18, 2026 in Ideas

  

  

  0