Hanzo AI
Try Hanzo

Benchmarks / Zen 6 / Zen 6 throughput & Latent Sparse Attention

Zen 6 · Throughput & subquadratic long-context attention

Zen 6 throughput & Latent Sparse Attention

Zen 6 (27.3B dense), the latest open-weight generation, measured on Hanzo hardware (NVIDIA Grace-Blackwell GB10, AMD Strix Halo, Apple silicon) and with Hierarchical Latent Sparse Attention (D-LSA / N-LSA) from 8K to 10M tokens.

Latest

42.50×D-LSA Decode speedup · Zen 6 · D-LSA (Tree) · 10M tokens (NVIDIA H200)Full Attention 1.00× · Win · Nov 4, 2026 · Hanzo-measured
78.57× (0.167 ms)D-LSA Decode speedup · Zen 6 · D-LSA (Tree) · 1M tokens (NVIDIA Grace-Blackwell GB10)Full Attention 1.00× (13.105 ms) · Win · Nov 4, 2026 · Hanzo-measured
42.75× (0.154 ms)D-LSA Decode speedup · Zen 6 · D-LSA (Tree) · 512K tokens (NVIDIA Grace-Blackwell GB10)Full Attention 1.00× (6.586 ms) · Win · Nov 4, 2026 · Hanzo-measured
23.35× (0.157 ms)D-LSA Decode speedup · Zen 6 · D-LSA (Tree) · 256K tokens (NVIDIA Grace-Blackwell GB10)Full Attention 1.00× (3.670 ms) · Win · Nov 4, 2026 · Hanzo-measured
72.40×D-LSA Prefill speedup · Zen 6 · D-LSA (Tree) · 10M tokens (NVIDIA H200)Full Attention 1.00× · Win · Nov 4, 2026 · Hanzo-measured
~70 tok/sDecode throughput · Zen 6 · NVIDIA Grace-Blackwellno baseline · Unscored · Nov 4, 2026 · Hanzo-measured
1.60× per layerN-LSA Training speedup · Zen 6 · N-LSA (Native) · 128K tokensFull Attention 1.00× · Win · Nov 4, 2026 · Hanzo-measured
~3,665 tok/sPrefill throughput · Zen 6 · NVIDIA Grace-Blackwellno baseline · Unscored · Nov 4, 2026 · Hanzo-measured

History

23 results across 5 metrics, newest first. Superseded runs stay, so a revision's progress can be read.

MeasuredSubjectMetricValueBaselineResultStatus
Nov 4, 2026Zen 6 · D-LSA (Tree) · 10M tokens (NVIDIA H200)D-LSA Decode speedup42.50×Full Attention 1.00×
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 1M tokens (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup78.57× (0.167 ms)Full Attention 1.00× (13.105 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 512K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup42.75× (0.154 ms)Full Attention 1.00× (6.586 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 256K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup23.35× (0.157 ms)Full Attention 1.00× (3.670 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA / N-LSA · 128K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup12.77× (0.152 ms)Full Attention 1.00× (1.944 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Flat) · 64K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup6.46× (0.111 ms)Full Attention 1.00× (0.716 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Flat) · 32K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup4.39× (0.135 ms)Full Attention 1.00× (0.593 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA GPU Cache · 8K tokens (Hanzo Engine live) (NVIDIA Grace-Blackwell GB10)D-LSA Decode speedup4.92× (70.0 tok/s)Full Attention 1.00× (14.2 tok/s)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 10M tokens (NVIDIA H200)D-LSA Prefill speedup72.40×Full Attention 1.00×
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 1M tokens (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup46.29× (26.40 ms)Full Attention 1.00× (1,222.10 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 512K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup35.75× (17.09 ms)Full Attention 1.00× (610.86 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Tree) · 256K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup27.50× (11.25 ms)Full Attention 1.00× (309.25 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA / N-LSA · 128K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup23.72× (6.46 ms)Full Attention 1.00× (153.29 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Flat) · 64K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup12.17× (6.71 ms)Full Attention 1.00× (81.62 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA (Flat) · 32K tokens (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup8.58× (4.28 ms)Full Attention 1.00× (36.70 ms)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · D-LSA GPU Cache · 8K tokens (Hanzo Engine live) (NVIDIA Grace-Blackwell GB10)D-LSA Prefill speedup5.53× (3,665.5 tok/s)Full Attention 1.00× (662.3 tok/s)
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · NVIDIA Grace-BlackwellDecode throughput~70 tok/s–
Unscored
Live · Hanzo-measured
Nov 4, 2026Zen 6 · AMD Strix HaloDecode throughput~37 tok/s–
Unscored
Live · Hanzo-measured
Nov 4, 2026Zen 6 · Apple siliconDecode throughput64–81 tok/s–
Unscored
Live · Hanzo-measured
Nov 4, 2026Zen 6 · N-LSA (Native) · 128K tokensN-LSA Training speedup1.60× per layerFull Attention 1.00×
Win
Live · Hanzo-measured
Nov 4, 2026Zen 6 · NVIDIA Grace-BlackwellPrefill throughput~3,665 tok/s–
Unscored
Live · Hanzo-measured
Nov 4, 2026Zen 6 · AMD Strix HaloPrefill throughput1,150–1,452 tok/s–
Unscored
Live · Hanzo-measured
Nov 4, 2026Zen 6 · Apple siliconPrefill throughput500–690 tok/s–
Unscored
Live · Hanzo-measured

Conditions

NVIDIA Grace-Blackwell: 121 GB unified, NVFP4 + Latent Sparse Attention (D-LSA) (141.2 tok/s with DFlash2 · 46.29× prefill & 78.57× decode at 1M context (99.8% KV pruned))AMD Strix Halo: 128 GB unified, 4-bit build (0.1 ms warm snapshot TTFT · 262K context in two slots)Apple silicon: unified memory, MLX buildD-LSA (Distilled Latent Sparse Attention): 512-token blocks, 4-key pooling, mutual-cheapest Ward clustering into 16 semantic buckets (bucket_size = 8, bucket_cap = 16), top_p = 128 buckets, top_k = 512 pooled keys, window_size = 512, sink_size = 4N-LSA (Native Latent Sparse Attention): trained from scratch on next-token cross-entropy loss alone; 1.6× faster per-layer training at 128K tokens while matching full-attention accuracydecode is one request unless a row says otherwisea range is what the runs spread over; the low end is the figure ordered and drawn

Hardware

NVIDIA Grace-Blackwell · AMD Strix Halo · Apple silicon · NVIDIA H200

Source

Zen 6: Subquadratic Ultra-Long Context via Hierarchical Latent Sparse Attention (Nov 4, 2026)

Reproduce

hanzo-server --port 36900 --no-paged-attn run -m zenlm/zen6 --max-seq-len 1048576

More Zen 6 benchmarks

Engine prefill on CUDA vs llama.cpp

Every result for Zen 6 throughput & Latent Sparse Attention on the shelf · How these are measured