# Hanzo — Benchmarks

> Independent AI benchmarks: Enso model routing, Zen reasoning, Kai decisions, memory retrieval latency, and engine throughput.

Hanzo AI · Benchmarks · Last measured Oct 16, 2026

# Intelligence, cost and speed, each measured.

Zen 6 prefills ~2,400 tokens a second and decodes ~43 on NVIDIA Grace-Blackwell, measured Sep 23, 2026.

[The results](#enso)[Every benchmark](#all)

[Enso](#enso)[Kai](#kai)[Zen 6](#zen)[Agent harness](#harness)[Agent platform](#platform)[Methods](#methods)[Every benchmark](#all)Jump to a benchmark/

Enso · intelligence per dollar

## Enso on the boards it was run on.

## The frontier, one benchmark at a time

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.

BenchmarkGPQA Diamond · 12 runs

LiveCodeBench · 12 runs

CSV

JSON

Copy link

One benchmark at a time: the frontier is drawn only among runs of the same benchmark, never across two. Filled points are Pareto-efficient — no other run scores higher for less. Hanzo's router is the large point in the brightest ink, labelled in bold; the single models are grey. A run's cost is the input and output tokens it records on the whole question set at the price list's rate for its own model id, the router's included, at list prices, Oct 31, 2026. Hover or tab to a point for its record.

The 12 runs as a table

System

Accuracy

Run cost

Tag, date

Pareto

enso

92.9%

$5.58

frontier · 2026-07-19

yes

gpt-5.6-sol

91.9%

$4.08

frontier · 2026-07-19

yes

opus-5.5

90.9%

$3.78

fill · 2026-10-16

yes

sonnet-5.5

90.4%

$1.85

fill · 2026-10-16

yes

gpt-5.5

89.9%

$11.95

frontier · 2026-07-19

fable-5.1

87.9%

$8.23

fill · 2026-10-16

opus-4.8

86.4%

$3.66

frontier · 2026-07-19

haiku-5.5

83.3%

$0.19

fill · 2026-10-16

yes

fable-5

82.8%

$9.87

frontier · 2026-07-19

claude-5-sonnet

77.8%

$5.66

frontier · 2026-07-19

deepseek-v4-pro

74.2%

$0.433

frontier · 2026-07-19

gpt-4o

54.5%

$1.58

frontier · 2026-07-19

## Every system on GPQA Diamond

GPQA Diamond · enso.gpqa_diamond, split test · every row a research run, read from api.hanzo.ai when the page opens · sorted by accuracy

System

Accuracy

Correct

Run cost

Tag, date

enso[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-enso)

92.9%

184 of 198

$5.58

frontier · 2026-07-19

enso-ultra[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-enso-ultra)

92.9%

184 of 198

$20.20

frontier · 2026-07-19

gpt-5.6-sol[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-gpt-5.6-sol)

91.9%

182 of 198

$4.08

frontier · 2026-07-19

opus-5.5[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-fill-opus-5.5)

90.9%

180 of 198

$3.78

fill · 2026-10-16

sonnet-5.5[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-fill-sonnet-5.5)

90.4%

179 of 198

$1.85

fill · 2026-10-16

gpt-5.5[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-gpt-5.5)

89.9%

178 of 198

$11.95

frontier · 2026-07-19

fable-5.1[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-fill-fable-5.1)

87.9%

174 of 198

$8.23

fill · 2026-10-16

opus-4.8[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-opus-4.8)

86.4%

171 of 198

$3.66

frontier · 2026-07-19

haiku-5.5[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-fill-haiku-5.5)

83.3%

165 of 198

$0.190

fill · 2026-10-16

fable-5[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-fable-5)

82.8%

164 of 198

$9.87

frontier · 2026-07-19

claude-5-sonnet[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-claude-5-sonnet)

77.8%

154 of 198

$5.66

frontier · 2026-07-19

deepseek-v4-pro[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-deepseek-v4-pro)

74.2%

147 of 198

$0.433

frontier · 2026-07-19

gpt-4o[run](https://api.hanzo.ai/v1/research/runs?org=hanzo&id=enso.gpqa_diamond-test-frontier-gpt-4o)

54.5%

108 of 198

$1.58

frontier · 2026-07-19

Each system's frontier run, else its v1 run, else its fill run: the three put the same prompts to a model. A run cost is the input and output tokens the run records at the price list's rate for its own model id, the router's included, at list prices, Oct 31, 2026; a dash is a run that records no tokens, a model the list does not price, or a list that could not be read.

Kai · decisions

## Kai beside Jev, on the same questions.

### Decision Suite

Reading the Decision Suite's research runs from api.hanzo.ai

## Kai against Jev, suite by suite

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.

Reading Kai's research runs from api.hanzo.ai

## 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.

Test

Kai

Jev

A question and its paraphrase, mean |Δp| over 20 pairs

0.124

0.151

A state silent on the question, answered between 0.35 and 0.65: share of 60

81.7%

38.3%

Accuracy over 951 validation rows of 9 yes/no suites, pooled

88.3%

87.5%

Kai is kai-1.2, read 2026-10-08; Jev is typesafe/jev-1.13 through the API, read 2026-10-06. Hanzo-measured; harness private.

## The trolley problem, decided on the counts

Kai does not weigh lives by preference. Its policy acts on the counts it grounds from the scene, and escalates to a person where it cannot ground them. On trolley-bench's 5,240 setups, scored by the article's own rules, the checks below are the ones Kai leads Jev on; every check is lower is better, and GPT-6 Luna's recorded answers stand beside both.

Check

Kai

Jev

GPT-6 Luna

Acted where acting kills more than it saves

17 of 672 (2.5%)

69 of 672 (10.3%)

143 of 672 (21.3%)

Killed a stranger to save the people, but would not destroy its own server

48 of 480 (10.0%)

180 of 480 (37.5%)

0 of 480 (0.0%)

Pulled the switch onto one person with nobody on the main track, of the wordings

0 of 24 (0.0%)

1 of 24 (4.2%)

24 of 24 (100.0%)

Kai is kai-1.8 through its nio/harm@1 policy, read 2026-10-20; Jev is typesafe/jev-1.13 and GPT-6 Luna is gpt-6-luna, each the benchmark's recorded answers. zachlandes/trolley-bench at cde97d7. Hanzo-scored by the article's rules; the scorer reproduces the article's numbers from the recorded answers.

Before any of it, Kai was qualified on 14,832 sealed trolley-structured scenes in two registers it had never read (quarry and piste), hashed before the model was frozen and read once. Every scene of the switch, server, robot, developer, and empty families, 8,300 in all, was decided right, with no wrong course; the crowd, hedged, infra, and vague families, where the harm cannot be counted, were escalated to a person every time.

Family

Scenes

Coverage

Accuracy

Wrong

switch

1,734

99.9%

100.0%

0

server

2,400

99.8%

100.0%

0

robot

1,366

99.9%

100.0%

0

developer

1,600

100.0%

100.0%

0

empty

1,200

100.0%

100.0%

0

spelled

552

98.6%

99.8%

1

bridge

300

58.0%

98.9%

2

ground-8, read 2026-10-20 at the qualified floor. Coverage is the share of scenes grounded; accuracy is of those; wrong counts the courses a grounded scene got wrong. Hanzo-measured; the sealed set is private.

## 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.

[Every request](#shortlist)

Recall@32, BM25

73.6

Recall@32, shortlist

93.6

nDCG@10, BM25

51.6

nDCG@10, shortlist

72.3

Top-1, BM25

41.4

Top-1, shortlist

55.5

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

[Requests for a Hanzo service](#shortlist)

Recall@32, BM25

15.5

Recall@32, shortlist

99.9

nDCG@10, BM25

3.5

nDCG@10, shortlist

81.5

Top-1, BM25

0.0

Top-1, shortlist

51.0

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

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

## Verified coding-agent performance

Hanzo Dev, with Kai and Enso Flash on api.hanzo.ai, against native Claude Code on held-out repository repairs. Each task is one fault its visible test fails on, and counts as solved only when a hidden verifier passes. Speed is the paired wall time over the tasks both solved.

[3.02×faster on the 42 tasks both solved95% interval 2.06–4.36×.](#coding-agent)Tasks both solved, by how Hanzo Dev settled them

Tasks

Speed

95% interval

Every task both solved

42

3.02×

2.06–4.36×

Settled by search, no model call

24

8.22×

6.86–9.82×

The model’s first edit

4

1.31×

1.14–1.48×

Agent

Median seconds

Model calls a task

Claude Code, sonnet-5.5 default effort

9.5

3.8

Hanzo Dev, dev flow fix, Kai + enso-flash

2.3

1.0

The fastest tasks are settled by a search with no model call: it tries single-token edits of the kinds these faults are drawn from against the visible test alone, and never reads a hidden test or the fault. These results are specific to this benchmark, one attempt a task; they are not a claim about general-purpose coding.

Research runs of coding.repair on split held, ended 2026-10-15, harness commit 8e966be: GET /v1/research/runs?benchmark=coding.repair.

## Agent tools

Tool handlers in Hanzo MCP, each run on the same cases as the handler it replaced. Each row says what decides it, rules, BM25 or Kai, as its run records. A rate is shown with its count.

### Shell commands an agent should not run unasked

Split hermes: hermes-jev-skills evals/gate/fixtures.jsonl (MIT), 118 made-up commands. The rules were frozen and read once (1 of 65 allowed: a metadata-endpoint fetch); that class and a bare `rm -rf /` were fixed after, so this run is not held out

Handler

What decides

Allowed, should deny or ask

Ordinary, stopped

Must deny, denied

Median time

Current

Rules

0.0% (0 / 65)

34.0% (18 / 53)

54.3% (19 / 35)

0.060 ms

Replaced, ae4d07b

Rules, then Kai

18.5% (12 / 65)

62.3% (33 / 53)

14.3% (5 / 35)

–

Research runs of mcp.command on split hermes, ended 2026-10-15; the current handler at commit 863cb8a: [GET /v1/research/runs?org=hanzo&benchmark=mcp.command](https://api.hanzo.ai/v1/research/runs?org=hanzo&benchmark=mcp.command)

### Which registered tool a task needs

Split t1dev: supplied tool lists from Kai's T1 tools6 dev split (manifest 48be841f)

Handler

What decides

Right tool first

MRR

Abstained

Median time

Current

Rules, then BM25

73.3% (220 / 300)

0.841

1

0.160 ms

Replaced, ae4d07b

Kai

65.7% (197 / 300)

0.805

43

–

Research runs of mcp.select on split t1dev, ended 2026-10-15; the current handler at commit 863cb8a: [GET /v1/research/runs?org=hanzo&benchmark=mcp.select](https://api.hanzo.ai/v1/research/runs?org=hanzo&benchmark=mcp.select)

Split hand: eight hand-labelled tasks over eight tools

Handler

What decides

Right tool first

MRR

Abstained

Median time

Current

Rules, then BM25

50.0% (4 / 8)

0.590

2

0.100 ms

Replaced, ae4d07b

Kai

25.0% (2 / 8)

0.491

2

–

Research runs of mcp.select on split hand, ended 2026-10-15; the current handler at commit 863cb8a: [GET /v1/research/runs?org=hanzo&benchmark=mcp.select](https://api.hanzo.ai/v1/research/runs?org=hanzo&benchmark=mcp.select)

### What kind of failure a CI log shows

Split ci60: 60 real CI logs from 48 public repositories, labelled from the excerpt alone without the rules in view; nameable = the 40 whose class the old six categories can name

Handler

What decides

Right class

Right class, nameable

Unknown

Median time

Current

Rules

70.0% (42 / 60)

70.0% (28 / 40)

9

0.370 ms

Replaced, ae4d07b

Kai

–

50.0% (20 / 40)

–

–

Research runs of mcp.triage on split ci60, ended 2026-10-15; the current handler at commit 863cb8a: [GET /v1/research/runs?org=hanzo&benchmark=mcp.triage](https://api.hanzo.ai/v1/research/runs?org=hanzo&benchmark=mcp.triage)

## Price

Reading Kai's price from api.hanzo.ai

Reading Kai's research runs from api.hanzo.ai

Zen 6 · speed on our own hardware

## Zen 6 prefills ~2,400 tokens a second on NVIDIA Grace-Blackwell.

[~2,400 tok/sPrefill · Zen 6 on NVIDIA Grace-BlackwellDecode ~43 tok/s · 50.6 tok/s decoding code · 112 tok/s across 8 requests · measured by Hanzo on Sep 23, 2026.](https://hanzo.ai/benchmarks/zen-6)[Prefill, tokens a second](https://hanzo.ai/benchmarks/zen-6)

NVIDIA Grace-Blackwell · NVFP4 weights~2,400 tok/s

AMD Strix Halo · 4-bit build1,150–1,400 tok/s

Apple silicon · MLX build500–690 tok/s

the low end of a range is the bar

[Decode, tokens a second](https://hanzo.ai/benchmarks/zen-6)

NVIDIA Grace-Blackwell · NVFP4 weights~43 tok/s

AMD Strix Halo · 4-bit build~34 tok/s

Apple silicon · MLX build64–81 tok/s

one request

Agent harness · what the agent remembers and finds

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

[2.0×Multi-hop ALL@20 · LoCoMo37.5% against cosine's 18.8% on the test split: every annotated evidence turn in the top 20.](https://hanzo.ai/benchmarks/locomo)

## 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](https://hanzo.ai/benchmarks/memoryagentbench)

single-hop, plain resolver

95.5

single-hop, typed search

98.3

multi-hop, plain resolver

56.8

multi-hop, typed search

84.5

262k multi-hop, plain

33.0

262k multi-hop, typed

67.0

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](https://hanzo.ai/benchmarks/locomo)

multi-hop ALL@20, cosine

18.8

multi-hop ALL@20, Hanzo

37.5

R@20, cosine

70.9

R@20, Hanzo

80.7

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

[LongMemEval](https://hanzo.ai/benchmarks/longmemeval)

cosine

85.8

engine, frozen on LoCoMo

90.6

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

## Knowledge graph

Point-in-time questions over facts that begin and end, against another temporal store on the same data. On the CronQuestions types the harness covers, the Hanzo graph answers 99.88% and Semantica 99.88%; what differs is the price of a question.

[0.063 msHanzo graph, through the APImedian question · Hits@1 99.88%](https://hanzo.ai/benchmarks/cronquestions)

[5,357 msSemantica temporal querymedian question · Hits@1 99.88%](https://hanzo.ai/benchmarks/cronquestions)

## Code

Cross-file retrieval in a repository, scored apart from generation. Typed links are regular expressions over imports, definitions and identifiers; no model reads anything in that row.

[RepoBench-R](https://hanzo.ai/benchmarks/repobench-r)

dense retrieval

18.8

typed links

31.4

cross-file-first · test split · R@1

## Agent Code Intelligence: The Race

How Hanzo Dev and Hanzo MCP accelerate agentic code comprehension compared to traditional raw-file reading and containerized local RAG stacks (Qdrant + Ollama). Benchmarked across production multi-language repositories.

Subsystem Logic (State Tracking & Auth)

Function Signature & Implementation

API Surface & Route Discovery

Claude Code Acceleration: Zero-Model Fault Repair (Kai)

Claude Code Acceleration: Atomic AST Refactoring

TARGET QUERY

Workspace state diffing & hashing logic

Run Race Simulation

⚡ HANZO NATIVEhanzo dev + hanzo-mcp

0.7 ms1 turn398 tokens (98.4% saved)

In-process Trigram + Tree-Sitter AST context bundle (0 daemons) · Zero Docker, Zero background model daemons

⚙️ LOCAL RAGQdrant Container + Ollama (phi3 / mxbai)

202 ms1 turn382 tokens

Qdrant Vector DB container + Ollama mxbai-embed-large (10 GB) · Requires 16 GB RAM + Docker daemon

🐌 TRADITIONAL AGENTClaude Code / Cursor (Raw Grep + File Reads)

36.4 s8 turns0 tokens

Iterative grep + full file reads over stdio (8 roundtrips to LLM)Burning prompt tokens...

Architecture

Latency

Tool Turns

Tokens Injected

Token Savings

Speedup

Traditional Agent (Grep + Whole File Reads)

36.4 s

8

25,078

Baseline

1.0×

Containerized Local RAG (Qdrant + Ollama)

202 ms

1

382

98.5%

180.0×

Hanzo Native (`hanzo dev` / `hanzo code claude`)

0.7 ms

1

398

98.4%

100×+

In a real agentic workflow (e.g. Claude Code, Cursor, Windsurf), each tool roundtrip incurs an additional 3–5 seconds of LLM inference latency. By replacing 6–13 iterative file exploration roundtrips with Hanzo’s in-process AST indexing (`code_context`, `code_tool`, `hanzo_refactor`) and Kai deterministic test-driven search, Hanzo Dev accelerates Claude Code by 20× to 50× (and up to 137× on zero-model search-settled faults), while slashing token consumption by 90% to 100%.

Research runs on benchmarks claude-acceleration and code-rag-navigation, harness commit hanzoai/benchmarks: [GET /v1/research/runs?org=hanzo&benchmark=claude-acceleration](https://api.hanzo.ai/v1/research/runs?org=hanzo&benchmark=claude-acceleration) · [GET /v1/research/runs?org=hanzo&benchmark=code-rag-navigation](https://api.hanzo.ai/v1/research/runs?org=hanzo&benchmark=code-rag-navigation) · [GET /v1/benchmark/leaderboard?benchmark=code-rag-navigation](https://api.hanzo.ai/v1/benchmark/leaderboard?benchmark=code-rag-navigation)

Agent platform · what an agent costs

## A dormant agent costs 477 bytes.

[477 bytesState per dormant agentAgainst a published ~1 MB (2,198×). 1M agents: 455 MB on disk vs ~977 GB.](https://hanzo.ai/benchmarks/platform-bench)

## Platform, agents, and sandboxes

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

[477 bytesState per dormant agent](https://hanzo.ai/benchmarks/fleet)

[601 bytesHeap per live agent, a goroutine](https://hanzo.ai/benchmarks/goroutine)

[125 nsSpawn one goroutine](https://hanzo.ai/benchmarks/goroutine)

[107 ms — 107 ns eachWake one million goroutines](https://hanzo.ai/benchmarks/goroutine)

A live agent here is a goroutine: it shares one address space with every other, and nothing about it is an isolation boundary. Code that has to be isolated gets a microVM. The goroutine timings are the median of one pass on an Apple M4 Max, 16 cores. Every figure here is transcribed from the harness's README rather than read from a run file, and each page names the command that prints it.

## Start-up and cost

[0.034 msResume one dormant agentIn process, from its stored state.](https://hanzo.ai/benchmarks/fleet)

[149.6 msA sandbox that runs a shell, coldpooled 35.8 ms; published [$0.00000063One call, marginal costAgainst a published $0.05 credit (79,614×).](https://hanzo.ai/benchmarks/platform-bench)

Methods

## How every number was taken

## Methodology

Every result is current. Each build pulls the latest validated runs from the research API: canonical, public, and not flagged inconsistent. When the API cannot be read, or answers with fewer runs than the committed snapshot already holds, the build keeps the snapshot, and the page swaps in what the API holds once it answers in the browser.

Every row is dated. A run is dated by when it ended; the platform bench records its month.

Live, frozen, superseded. Live is the newest run the API serves for a benchmark. Frozen is a committed run file, a snapshot, or a figure transcribed from a harness’s README. Superseded is an earlier run of the same benchmark that a later one replaced; it stays in the library, so a revision’s history can be read.

Where a number comes from. Kai’s and Enso’s rows are research runs, served at [the research API](https://api.hanzo.ai/v1/research/runs?org=hanzo) and read when this page opens. The memory, graph and code panels read the committed run files of the benchmarks repository. Kai’s yes/no table and the tool shortlist are Hanzo-measured reads the research API does not file: each is its run’s report, copied unchanged, and names its model, date and count.

Measured, not reported. Every frontier point is a run Hanzo filed: the router and each single model on the same questions, at the same tag. No vendor-reported figure is on a board here.

Prices are prices. Kai beside Jev is the input-token rate each is billed, from the price list, not a cost per task. The frontier’s axis is what each run spent on the whole question set. No speed is claimed here without a measured comparison; latency, where this Kai version was timed, is in its tables.

Comparing rows. A frontier is drawn among the rows of one benchmark, never across two, and a leaderboard row compares two models on one suite. Suites measure different tasks, so the mean is a summary, not a score on any of them.

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](https://hips.hanzo.ai).

[Kai: Calibrated Discrete Decision Intelligence for Autonomous Agent Loops](https://papers.hanzo.ai/kai-decision-model/)

papers.hanzo.aiFactorized Discrete Choice, Calibrated Distributions, and Deterministic Policy Governance

[Linkage, Not Similarity](https://papers.hanzo.ai/hanzo-context-memory/)

papers.hanzo.aiWhat long-term agent memory has to retrieve, and how it is measured

[Hanzo AI Chain (AIVM)](https://papers.hanzo.ai/hanzo-ai-chain/)

papers.hanzo.aiUseful-Work Mining, the Native AI Coin, and Post-Quantum Omnichain Settlement

[Cloud Economics of a Sovereign OSS Stack](https://papers.hanzo.ai/hanzo-cloud-economics/)

papers.hanzo.aiA Regulated Capital-Markets Migration Case Study

[One Native Stack for Private, Continuously-Learning AI](https://papers.hanzo.ai/hanzo-native-stack-thesis/)

papers.hanzo.aiInference at the Bandwidth Wall, On-Device QLoRA, One Engine

[Native ROCm Inference on a Consumer RDNA3.5 APU](https://papers.hanzo.ai/hanzo-rocm-inference/)

papers.hanzo.aiReaching llama.cpp Decode Parity via a Unified 1-bit-to-Full Quant Core

[Edge LLM Inference Across Commodity Accelerators](https://papers.hanzo.ai/hanzo-cross-backend-inference/)

papers.hanzo.aiMeasured AMD, NVIDIA, and Apple Performance

[Native Training in the Hanzo Engine](https://papers.hanzo.ai/hanzo-native-training/)

papers.hanzo.aiOne Engine, One Quant Format, One Backend

[Continuously-Learning Private AI](https://papers.hanzo.ai/hanzo-continuous-learning-privacy/)

papers.hanzo.aiSelf-Improvement and Privacy in the Hanzo Native Stack

[Economical Serving and the hanzo.ai Platform](https://papers.hanzo.ai/hanzo-platform-infra-desktop/)

papers.hanzo.aiA Bandwidth-Optimal Unified Train-and-Serve Runtime

[The Unified Sovereign-Tenant Cloud](https://papers.hanzo.ai/hanzo-unified-tenant-cloud/)

papers.hanzo.aiLinear Shared-Nothing Scaling via Per-Tenant SQLite and In-Process Composition

[Refutation-Driven Performance Engineering](https://papers.hanzo.ai/refutation-driven-perf/)

papers.hanzo.aiAn Empirical Study of a Multi-Week GPU Kernel Campaign

[Evolutionary Schedule Search in a One-Source Kernel DSL](https://papers.hanzo.ai/evolutionary-schedule-search/)

papers.hanzo.aiAnd the Continual-Improvement Loop It Anchors

[Hanzo Router](https://papers.hanzo.ai/hanzo-router/)

papers.hanzo.aiMemory-Aware, Local-First LLM Routing with a Learned SLO-Constrained Policy

[What a Dormant Agent Costs](https://papers.hanzo.ai/hanzo-dormant-agents/)

papers.hanzo.aiState, Execution and Isolation Measured Separately at a Fleet of One Million

[Hops Do Not Fix Multi-Hop Retrieval](https://papers.hanzo.ai/hanzo-multi-hop-retrieval/)

papers.hanzo.aiA Negative Result on Long-Horizon Conversational Memory, and What the Failure Actually Is

[Retrieval Scoped by Subject](https://papers.hanzo.ai/locomo-subject-scope/)

papers.hanzo.aiA factorial measurement of conversational memory on LoCoMo, and the negative controls that locate the effect

[Hanzo Network Whitepaper](https://papers.hanzo.ai/network-whitepaper/)

papers.hanzo.aiL1 Blockchain for Decentralized AI Compute

Every benchmark

## Every benchmark, every run

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.

OrderNewest

Largest lead

Benchmark

Value

CSV

JSON

Copy link

Views[Live results](?state=live#all)[Kai, every revision](?product=Kai&category=Decision%20harness&measure=Accuracy&order=name#all)[Superseded runs](?state=superseded#all)[Zen 6 and the engine](?product=Zen%206#all)[Not measured by Hanzo](?by=Vendor&by=Board#all)

Filters

ProductKai120

Zen 611

Agent harness8

Agent platform12

CategoryDecision harness115

Platform, agents, and sandboxes9

Throughput on Hanzo hardware8

Memory and retrieval6

Capability3

Cost and start-up3

Engine prefill vs llama.cpp3

Latency2

2 more

MetricAccuracy67

Calibration error (ECE)51

Decode throughput3

Prefill throughput3

Prefill throughput vs llama.cpp3

substring EM3

Also measured2

A dormant agent1

18 more

Measured byHanzo151

DateOct 202642

Sep 2026106

Jul 20263

StatusLive25

Frozen31

Superseded95

ResultWin129

Unscored22

151 of 151 results, 24 benchmarks

[AG News](https://hanzo.ai/benchmarks/kai-ag-news)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

93.8%baseline 86.0%

Win

Oct 16, 2026Live · Hanzo-measured

[AG News](https://hanzo.ai/benchmarks/kai-ag-news)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.022baseline 0.102

Win

Oct 16, 2026Live · Hanzo-measured

[Banking77](https://hanzo.ai/benchmarks/kai-banking77)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

88.0%baseline 83.5%

Win

Oct 16, 2026Live · Hanzo-measured

[Choosing among many options](https://hanzo.ai/benchmarks/kai-choice-size)kai-1.4 · c1856a77 · 4 options vs JevKai · Capability · Accuracy

94.3%baseline 90.5%

Win

Oct 16, 2026Live · Hanzo-measured

[Choosing among many options](https://hanzo.ai/benchmarks/kai-choice-size)kai-1.4 · c1856a77 · 16 options vs JevKai · Capability · Accuracy

92.0%baseline 85.3%

Win

Oct 16, 2026Live · Hanzo-measured

[Choosing among many options](https://hanzo.ai/benchmarks/kai-choice-size)kai-1.4 · c1856a77 · 77 options vs JevKai · Capability · Accuracy

88.0%baseline 84.3%

Win

Oct 16, 2026Live · Hanzo-measured

[DAIR Emotion](https://hanzo.ai/benchmarks/kai-dair-emotion)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

92.3%baseline 60.3%

Win

Oct 16, 2026Live · Hanzo-measured

[DAIR Emotion](https://hanzo.ai/benchmarks/kai-dair-emotion)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.033baseline 0.280

Win

Oct 16, 2026Live · Hanzo-measured

[Email spam](https://hanzo.ai/benchmarks/kai-email-spam)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

98.5%baseline 97.8%

Win

Oct 16, 2026Live · Hanzo-measured

[Email spam](https://hanzo.ai/benchmarks/kai-email-spam)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.016baseline 0.060

Win

Oct 16, 2026Live · Hanzo-measured

[MASSIVE](https://hanzo.ai/benchmarks/kai-massive)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

90.7%baseline 89.0%

Win

Oct 16, 2026Live · Hanzo-measured

[MASSIVE](https://hanzo.ai/benchmarks/kai-massive)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.064baseline 0.064

Win

Oct 16, 2026Live · Hanzo-measured

[Model routing](https://hanzo.ai/benchmarks/kai-model-routing)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

99.8%baseline 97.7%

Win

Oct 16, 2026Live · Hanzo-measured

[Model routing](https://hanzo.ai/benchmarks/kai-model-routing)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.002baseline 0.014

Win

Oct 16, 2026Live · Hanzo-measured

[Phishing](https://hanzo.ai/benchmarks/kai-phishing)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

98.8%baseline 90.0%

Win

Oct 16, 2026Live · Hanzo-measured

[Phishing](https://hanzo.ai/benchmarks/kai-phishing)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.011baseline 0.039

Win

Oct 16, 2026Live · Hanzo-measured

[RAG relevance](https://hanzo.ai/benchmarks/kai-rag-relevance)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

65.3%baseline 62.0%

Win

Oct 16, 2026Live · Hanzo-measured

[RAG relevance](https://hanzo.ai/benchmarks/kai-rag-relevance)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.089baseline 0.279

Win

Oct 16, 2026Live · Hanzo-measured

[Support triage](https://hanzo.ai/benchmarks/kai-support-triage)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

46.3%baseline 36.5%

Win

Oct 16, 2026Live · Hanzo-measured

[Support triage](https://hanzo.ai/benchmarks/kai-support-triage)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.349baseline 0.483

Win

Oct 16, 2026Live · Hanzo-measured

[Toxicity](https://hanzo.ai/benchmarks/kai-toxicity)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

81.0%baseline 66.3%

Win

Oct 16, 2026Live · Hanzo-measured

[Toxicity](https://hanzo.ai/benchmarks/kai-toxicity)kai-1.4 · c1856a77 vs JevKai · Decision harness · Calibration error (ECE)

0.170baseline 0.177

Win

Oct 16, 2026Live · Hanzo-measured

[Typed decisions](https://hanzo.ai/benchmarks/kai-typed-decisions)kai-1.4 · c1856a77 vs JevKai · Decision harness · Accuracy

74.6%baseline 73.6%

Win

Oct 16, 2026Live · Hanzo-measured

[AG News](https://hanzo.ai/benchmarks/kai-ag-news)kai-1 · 0834a74f vs JevKai · Decision harness · Accuracy

94.3%baseline 86.0%

Win

Oct 3, 2026Superseded · Hanzo-measured

Show all 151

## Every benchmark page

[Enso](?product=Enso#all)

[GPQA-Diamond](https://hanzo.ai/benchmarks/gpqa)filter the shelf to it

[Kai](?product=Kai#all)

[AG News](https://hanzo.ai/benchmarks/kai-ag-news)[DAIR Emotion](https://hanzo.ai/benchmarks/kai-dair-emotion)[Banking77](https://hanzo.ai/benchmarks/kai-banking77)[Support triage](https://hanzo.ai/benchmarks/kai-support-triage)[Email spam](https://hanzo.ai/benchmarks/kai-email-spam)[Phishing](https://hanzo.ai/benchmarks/kai-phishing)[Toxicity](https://hanzo.ai/benchmarks/kai-toxicity)[RAG relevance](https://hanzo.ai/benchmarks/kai-rag-relevance)[Model routing](https://hanzo.ai/benchmarks/kai-model-routing)[Typed decisions](https://hanzo.ai/benchmarks/kai-typed-decisions)[MASSIVE](https://hanzo.ai/benchmarks/kai-massive)[Choosing among many options](https://hanzo.ai/benchmarks/kai-choice-size)[Latency of one decision](https://hanzo.ai/benchmarks/kai-latency)filter the shelf to it

[Zen 6](?product=Zen%206#all)

[Zen 6 throughput](https://hanzo.ai/benchmarks/zen-6)[Engine prefill on CUDA vs llama.cpp](https://hanzo.ai/benchmarks/inference)filter the shelf to it

[Agent harness](?product=Agent%20harness#all)

[LoCoMo](https://hanzo.ai/benchmarks/locomo)[MemoryAgentBench](https://hanzo.ai/benchmarks/memoryagentbench)[LongMemEval](https://hanzo.ai/benchmarks/longmemeval)[CronQuestions](https://hanzo.ai/benchmarks/cronquestions)[RepoBench-R](https://hanzo.ai/benchmarks/repobench-r)[LoCoMo · subject scope](https://hanzo.ai/benchmarks/locomo-subject-scope)[LoCoMo-Conv](https://hanzo.ai/benchmarks/locomo-conv)filter the shelf to it

[Agent platform](?product=Agent%20platform#all)

[Platform bench](https://hanzo.ai/benchmarks/platform-bench)[Fleet residency](https://hanzo.ai/benchmarks/fleet)[Live agent footprint](https://hanzo.ai/benchmarks/goroutine)[Sandbox cold start](https://hanzo.ai/benchmarks/sandbox)filter the shelf to it

Hanzo AI Cloud

### Build on Hanzo Cloud

[Start on Hanzo Cloud](https://cloud.hanzo.ai)[View Models & Pricing](https://hanzo.ai/pricing)