fix: restore MiniCPM4-8B in hackathon narrative, fix fine-tune duration claim

blog/index.html

@@ -553,9 +553,9 @@ footer { position: relative; z-index: 1; max-width: 1060px; margin: 0 auto; padd
 
   <h3 class="reveal" id="lesson-1">Model choice is a compatibility problem, not a benchmark problem</h3>
 
-  <p class="reveal">We started with MiniCPM3-4B, upgraded to MiniCPM4.1-8B for better reasoning, and immediately hit the classic open-model trap: the model card says one thing, the <code>transformers</code> version on your machine says another.</p>
+  <p class="reveal">We started with MiniCPM3-4B, upgraded to MiniCPM4-8B for better reasoning, and immediately hit the classic open-model trap: the model card says one thing, the <code>transformers</code> version on your machine says another.</p>
 
-  <span class="commit reveal">fix: pin transformers==4.57.1 for MiniCPM4.1-8B compatibility</span>
+  <span class="commit reveal">fix: pin transformers==4.57.1 for MiniCPM4-8B compatibility</span>
 
   <p class="reveal">The follow-up lesson came from quantization. Bitsandbytes 4-bit is great on a 16 GB local GPU and completely unnecessary on ZeroGPU hardware, so we made it conditional:</p>
 
@@ -687,7 +687,7 @@ footer { position: relative; z-index: 1; max-width: 1060px; margin: 0 auto; padd
     <div class="card"><div class="glare"></div><span class="idx">TRACE</span><h4>Agent trace on the Hub</h4><p>12 live LLM calls across 3 sessions (OS, ML, Networking) — exact prompts, raw outputs, timings — published as a dataset at <code>build-small-hackathon/PaperProf-traces</code> for the community to learn from.</p></div>
   </div>
 
-  <p class="reveal">The upgrade took less than an hour of code changes. The fine-tune ran overnight on a Modal A100-80GB. The lesson from the hackathon held: <strong>constraints make the architecture honest</strong>, and a well-structured pipeline makes iteration cheap.</p>
+  <p class="reveal">The upgrade took less than an hour of code changes. The fine-tune ran in ~20 minutes on a Modal A100-80GB. The lesson from the hackathon held: <strong>constraints make the architecture honest</strong>, and a well-structured pipeline makes iteration cheap.</p>
 
 </main>