Skip to content
Try Hanzo
Language: English (United States)

Intelligence, cost and speed, each measured.

Zen 6 prefills ~3,665 tokens a second and decodes ~70 on NVIDIA Grace-Blackwell, measured Nov 4, 2026.

Hanzo AI · Benchmarks · Last measured Oct 16, 2026

Zen 6 · speed on our own hardware & Latent Sparse Attention

NVIDIA Grace-Blackwell · NVFP4 + Latent Sparse Attention (D-LSA)~3,665 tok/s
AMD Strix Halo · 4-bit build1,150–1,452 tok/s
Apple silicon · MLX build500–690 tok/s

the low end of a range is the bar

~3,665 tok/s

Zen 6 · speed on our own hardware & Latent Sparse Attention

Zen 6 prefills ~3,665 tokens a second on NVIDIA Grace-Blackwell and scales 72.40× faster at 10M tokens with Latent Sparse Attention.

Prefill · Zen 6 on NVIDIA Grace-BlackwellDecode ~70 tok/s · 141.2 tok/s with DFlash2 · 46.29× prefill & 78.57× decode at 1M context (99.8% KV pruned) · measured by Hanzo on Nov 4, 2026.

Prefill, tokens a second

  • NVIDIA Grace-Blackwell · NVFP4 + Latent Sparse Attention (D-LSA)~3,665 tok/s
  • AMD Strix Halo · 4-bit build1,150–1,452 tok/s
  • Apple silicon · MLX build500–690 tok/s

the low end of a range is the bar

Decode, tokens a second

  • NVIDIA Grace-Blackwell · NVFP4 + Latent Sparse Attention (D-LSA)~70 tok/s
  • AMD Strix Halo · 4-bit build~37 tok/s
  • Apple silicon · MLX build64–81 tok/s

one request

A New Pareto Frontier of Intelligence

GPQA Diamond challenges AI with 198 graduate-level science questions. See how Enso balances accuracy and cost across 1,000 answers.

Ensosingle model, same 198 questionsreported:GLM 5.3, Grok 4.7, Kimi K3, GPT-6 Sol

GPQA-Diamond

7580859095100$1$3$10$30cheaper ← $ per 1,000 questions → dearerGPQA Diamond %GLM 5.3Enso FlashEnsoEnso Ultra

Enso on the boards it was run on.

Enso · intelligence per dollar

Accuracy against what the run cost at list prices, on a log axis, for each board where Enso’s router is on the frontier. Every point is a research run, read when the page opens, and its cost is the tokens it records on the whole question set at the price list’s rate for its own model id. The dashed line joins the runs no other run beats on both.

The frontier, one benchmark at a timeBenchmark
LiveCodeBench · 12 runs ↗
50556065707580859095$0.3$1$3$10LIST-PRICE COST OF THE WHOLE SET, $, LOG SCALEhaiku-5.5sonnet-5.5opus-5.5gpt-5.6-solenso

Kai wins 16 of 16 tasks of the Decision Suite against Jev.

Kai · decisions

Each task was fixed on 2026-10-15. The frozen harness's suites Kai leads are below.

Kai against Jev, suite by suite

Leaderboard

The suites Kai leads, both models on the same questions, each row dated by its runs. Sort by the lead, switch to calibration error, draw it as a chart or a table; the view is kept in the address, and the rows are read from the research API when the page opens.

11

Suites won · Kai vs Jev AG News, DAIR Emotion, Banking77, Support triage, Email spam, Phishing, Toxicity, RAG relevance, Model routing, Typed decisions, MASSIVE, 51 languages. Every row a research run of kai-1.4 and of Jev on the same questions; mean accuracy across these 11 suites 84.4% against 76.6%.

Metric
View
Sort

11 suites won · mean accuracy across these 11 suites 84.4% against 76.6%

  1. 01DAIR EmotionKai Oct 16, 2026 · Jev Sep 25, 2026Kai92.3%Jev60.3%+32.0 ppWin
  2. 02ToxicityKai Oct 16, 2026 · Jev Sep 25, 2026Kai81.0%Jev66.3%+14.8 ppWin
  3. 03Support triageKai Oct 16, 2026 · Jev Sep 25, 2026Kai46.3%Jev36.5%+9.8 ppWin
  4. 04PhishingKai Oct 16, 2026 · Jev Sep 25, 2026Kai98.8%Jev90.0%+8.8 ppWin
  5. 05AG NewsKai Oct 16, 2026 · Jev Sep 25, 2026Kai93.8%Jev86.0%+7.8 ppWin
  6. 06Banking77Kai Oct 16, 2026 · Jev Sep 25, 2026Kai88.0%Jev83.5%+4.5 ppWin
  7. 07RAG relevanceKai Oct 16, 2026 · Jev Sep 25, 2026Kai65.3%Jev62.0%+3.2 ppWin
  8. 08Model routingKai Oct 16, 2026 · Jev Sep 25, 2026Kai99.8%Jev97.7%+2.0 ppWin
  9. 09MASSIVE, 51 languagesKai Oct 16, 2026 · Jev Sep 25, 2026Kai90.7%Jev89.0%+1.7 ppWin
  10. 10Typed decisionsKai Oct 16, 2026 · Jev Sep 25, 2026Kai74.6%Jev73.6%+1.0 ppWin
  11. 11Email spamKai Oct 16, 2026 · Jev Sep 25, 2026Kai98.5%Jev97.8%+0.7 ppWin
  • Every cell a research run of kai-1.4 and of Jev, on the same questions.
  • Accuracy: higher is better; a lead is in percentage points.
  • Suites measure different tasks, so their rows are compared within a row, not across.

Yes or no, read from the state

Production Kai answers a yes/no question with a reader trained for it. The tests below are the ones it leads Jev on, both models given the same questions and the same rows.

  • A question and its paraphrase, mean |Δp| over 20 pairsKai0.124Jev0.151
  • A state silent on the question, answered between 0.35 and 0.65: share of 60Kai81.7%Jev38.3%
  • Accuracy over 951 validation rows of 9 yes/no suites, pooledKai88.3%Jev87.5%

Finding the right tool in a large catalog

Kai’s tool-relevance head, in development, ranks a shortlist of 32 tools drawn from the whole catalog. The shortlist is an open embedding model, zen-embedding-0.6B, with a small BM25 term; nothing is trained for it. Here it is scored against BM25 alone on a dev catalog Hanzo assembled from public tool-calling sets and its own API: 2,928 tools, 6,608 requests.

An internal dev set, not a public benchmark: one run, 2026-10-11. Retrieval only, not an agent result.

Every request

  • Recall@32, BM2573.6Recall@32, shortlist93.6
  • nDCG@10, BM2551.6nDCG@10, shortlist72.3
  • Top-1, BM2541.4Top-1, shortlist55.5

6,608 requests · 2,928 tools · Recall@32: a right tool among the first 32

Requests for a Hanzo service

  • Recall@32, BM2515.5Recall@32, shortlist99.9
  • nDCG@10, BM253.5nDCG@10, shortlist81.5
  • Top-1, BM250.0Top-1, shortlist51.0

670 requests · a service scored as the best of its operations

On LoCoMo, Hanzo's context engine finds 2.0× the multi-hop evidence cosine does.

Agent harness · what the agent remembers and finds

Memory and retrieval

What a memory finds in a long conversation, and what a reader does with it. Each panel is one metric, the baseline and Hanzo on the same data, the same embedder and the same cut-off, drawn on one scale.

MemoryAgentBench

  • single-hop, plain resolver95.5single-hop, typed search98.3
  • multi-hop, plain resolver56.8multi-hop, typed search84.5
  • 262k multi-hop, plain33.0262k multi-hop, typed67.0

substring EM · pooled over the dev haystack and the three test haystacks, and the largest alone · reader: none, embedder zenlm/zen-embedding-0.6b

LoCoMo

  • multi-hop ALL@20, cosine18.8multi-hop ALL@20, Hanzo37.5
  • R@20, cosine70.9R@20, Hanzo80.7

test split · zen-embedding-0.6b · k=20 · ALL = every annotated evidence turn in the top 20

LongMemEval

  • cosine85.8engine, frozen on LoCoMo90.6

500 questions · all-minilm vectors for both · ALL@5 = every evidence session in the top 5

Platform, agents, and sandboxes

What an agent costs at rest, what a live one costs to wake, and what a sandbox costs to start.

How every number was taken

Methods

The memory and code panels regenerate from the committed runs on a fresh clone:

git clone https://github.com/hanzoai/benchmarks && cd benchmarks/brain && node results.mjs

The research

Every paper the library publishes. A benchmark’s own page names the paper that reports it, where one does. Specifications are the HIPs.

Every benchmark, every run

Every benchmark

One shelf for everything above, by the same five products: each Kai revision on each suite it leads, each Enso tier on the frontiers it holds, each Zen 6 reading, each memory and platform pass. Search it, filter it, pick rows of one benchmark to compare, and open any of them for its history, conditions, hardware, source and the command that reproduces it.

Order

24 of 166 results, 24 benchmarks

  1. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 10M tokens (NVIDIA H200) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup72.40×baseline 1.00×WinNov 4, 2026Live · Hanzo-measured
  2. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 10M tokens (NVIDIA H200) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup42.50×baseline 1.00×WinNov 4, 2026Live · Hanzo-measured
  3. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 1M tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup46.29× (26.40 ms)baseline 1.00× (1,222.10 ms)WinNov 4, 2026Live · Hanzo-measured
  4. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 1M tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup78.57× (0.167 ms)baseline 1.00× (13.105 ms)WinNov 4, 2026Live · Hanzo-measured
  5. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 512K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup35.75× (17.09 ms)baseline 1.00× (610.86 ms)WinNov 4, 2026Live · Hanzo-measured
  6. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 512K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup42.75× (0.154 ms)baseline 1.00× (6.586 ms)WinNov 4, 2026Live · Hanzo-measured
  7. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 256K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup27.50× (11.25 ms)baseline 1.00× (309.25 ms)WinNov 4, 2026Live · Hanzo-measured
  8. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Tree) · 256K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup23.35× (0.157 ms)baseline 1.00× (3.670 ms)WinNov 4, 2026Live · Hanzo-measured
  9. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA / N-LSA · 128K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup23.72× (6.46 ms)baseline 1.00× (153.29 ms)WinNov 4, 2026Live · Hanzo-measured
  10. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA / N-LSA · 128K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup12.77× (0.152 ms)baseline 1.00× (1.944 ms)WinNov 4, 2026Live · Hanzo-measured
  11. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Flat) · 64K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup12.17× (6.71 ms)baseline 1.00× (81.62 ms)WinNov 4, 2026Live · Hanzo-measured
  12. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Flat) · 64K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup6.46× (0.111 ms)baseline 1.00× (0.716 ms)WinNov 4, 2026Live · Hanzo-measured
  13. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Flat) · 32K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup8.58× (4.28 ms)baseline 1.00× (36.70 ms)WinNov 4, 2026Live · Hanzo-measured
  14. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA (Flat) · 32K tokens (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup4.39× (0.135 ms)baseline 1.00× (0.593 ms)WinNov 4, 2026Live · Hanzo-measured
  15. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA GPU Cache · 8K tokens (Hanzo Engine live) (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Prefill speedup5.53× (3,665.5 tok/s)baseline 1.00× (662.3 tok/s)WinNov 4, 2026Live · Hanzo-measured
  16. Zen 6 throughput & Latent Sparse AttentionZen 6 · D-LSA GPU Cache · 8K tokens (Hanzo Engine live) (NVIDIA Grace-Blackwell GB10) vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · D-LSA Decode speedup4.92× (70.0 tok/s)baseline 1.00× (14.2 tok/s)WinNov 4, 2026Live · Hanzo-measured
  17. Zen 6 throughput & Latent Sparse AttentionZen 6 · N-LSA (Native) · 128K tokens vs Full AttentionZen 6 · Throughput & subquadratic long-context attention · N-LSA Training speedup1.60× per layerbaseline 1.00×WinNov 4, 2026Live · Hanzo-measured
  18. Zen 6 throughput & Latent Sparse AttentionZen 6 · NVIDIA Grace-BlackwellZen 6 · Throughput & subquadratic long-context attention · Prefill throughput~3,665 tok/sno baselineUnscoredNov 4, 2026Live · Hanzo-measured
  19. Zen 6 throughput & Latent Sparse AttentionZen 6 · NVIDIA Grace-BlackwellZen 6 · Throughput & subquadratic long-context attention · Decode throughput~70 tok/sno baselineUnscoredNov 4, 2026Live · Hanzo-measured
  20. Zen 6 throughput & Latent Sparse AttentionZen 6 · AMD Strix HaloZen 6 · Throughput & subquadratic long-context attention · Prefill throughput1,150–1,452 tok/sno baselineUnscoredNov 4, 2026Live · Hanzo-measured
  21. Zen 6 throughput & Latent Sparse AttentionZen 6 · AMD Strix HaloZen 6 · Throughput & subquadratic long-context attention · Decode throughput~37 tok/sno baselineUnscoredNov 4, 2026Live · Hanzo-measured
  22. Zen 6 throughput & Latent Sparse AttentionZen 6 · Apple siliconZen 6 · Throughput & subquadratic long-context attention · Prefill throughput500–690 tok/sno baselineUnscoredNov 4, 2026Live · Hanzo-measured
  23. Zen 6 throughput & Latent Sparse AttentionZen 6 · Apple siliconZen 6 · Throughput & subquadratic long-context attention · Decode throughput64–81 tok/sno baselineUnscoredNov 4, 2026Live · Hanzo-measured
  24. AG Newskai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy93.8%baseline 86.0%WinOct 16, 2026Live · Hanzo-measured

Build on Hanzo Cloud