docs(benchmarks): correct Gemma 3 1B to 241 tok/s at 256 tokens

Re-verified at standard 256 token count: 241 tok/s Zerfoo vs 201 tok/s
Ollama (1.20x, 20% advantage confirmed). The earlier 236 tok/s was
measured at 128 tokens where CUDA graph amortization is lower.

Author: dndungu

content/docs/reference/benchmarks.md

@@ -36,12 +36,13 @@ All official benchmarks run on a single machine:
 
 ### Multi-Model Comparison: Zerfoo vs Ollama (2026-03-25)
 
-Head-to-head decode throughput on DGX Spark GB10. 128 tokens, 3 runs (median),
-greedy sampling (temp=0), commit `294aa43` (v1.19.0), Ollama v0.17.7.
+Head-to-head decode throughput on DGX Spark GB10. 128 tokens (except where
+noted), 3 runs (median), greedy sampling (temp=0), commit `294aa43` (v1.19.0),
+Ollama v0.17.7.
 
 | Model | Architecture | Size | Zerfoo (tok/s) | Ollama (tok/s) | Ratio | Winner |
 |-------|-------------|------|----------------|----------------|-------|--------|
-| Gemma 3 1B Q4_K_M | gemma3 | 1B | **236.38** | 204.37 | **1.16x** | Zerfoo |
+| Gemma 3 1B Q4_K_M | gemma3 | 1B | **241** (256 tok) | 201 (256 tok) | **1.20x** | Zerfoo |
 | DeepSeek R1 1.5B Q4_K_M | deepseek2 | 1.5B | **192.83** | 184.75 | **1.04x** | Zerfoo |
 | Llama 3.2 3B Q4_K_M | llama | 3B | 96.06 | 97.66 | 0.98x | ~Even |
 | Mistral 7B Q4_K_M | mistral | 7B | 11.61 | 46.77 | 0.25x | Ollama |
@@ -105,16 +106,16 @@ dp4a benefits will appear at larger batch sizes where compute becomes the bottle
 
 | Framework | Version | Tokens | Tok/s (decode) | CUDA Graphs | Notes |
 |-----------|---------|--------|----------------|-------------|-------|
-| **Zerfoo** | v1.19.0 | 128 | **236.38** | Yes | Multi-model benchmark (2026-03-25) |
+| **Zerfoo** | v1.19.0 | 256 | **241** | Yes | Multi-model benchmark (2026-03-25) |
 | **Zerfoo** | v0.x | 256 | **244.45** | Yes | Single-model baseline (2026-03-20) |
 | **Zerfoo** | v0.x | 256 | 174.44 | No | Without CUDA graph capture |
 | **Ollama** | 0.17.7 | 128 | 204.37 | N/A | Multi-model benchmark (2026-03-25) |
 | **llama.cpp** | b5220+ | 256 | ~210-230 | No | Estimated from community reports on GB10-class hardware |
 
 **Summary:**
 
-- Zerfoo with CUDA graphs: **236 tok/s** (+16% vs Ollama, ~5-10% vs llama.cpp)
-- Zerfoo without CUDA graphs: **174 tok/s** (CUDA graph capture adds +36%)
+- Zerfoo with CUDA graphs: **241 tok/s** (+20% vs Ollama, ~5-15% vs llama.cpp)
+- Zerfoo without CUDA graphs: **174 tok/s** (CUDA graph capture adds +38%)
 - Ollama: **204 tok/s** (uses llama.cpp under the hood with its own overhead)
 
 > **Note on llama.cpp numbers:** Direct llama.cpp measurements on this exact
@@ -149,7 +150,7 @@ dp4a benefits will appear at larger batch sizes where compute becomes the bottle
 
 | GPU | Zerfoo (est.) | Notes |
 |-----|---------------|-------|
-| DGX Spark GB10 | 236 tok/s | Measured (Gemma 3 1B, 2026-03-25) |
+| DGX Spark GB10 | 241 tok/s | Measured (Gemma 3 1B, 2026-03-25) |
 | RTX 4090 | TBD | Community contributions welcome |
 | RTX 3090 | TBD | Community contributions welcome |
 | A100 80GB | TBD | Community contributions welcome |