Try Hanzo
SYSTEMS BENCHMARKS

98.0% on GPQA-Diamond, and What It Costs

Tiered dispatch reaches frontier accuracy at a fraction of frontier price. Below it: what an agent costs at rest and in flight, how fast a sandbox starts, and what retrieval recovers — with the losses and the gaps.

01   MODEL ROUTING

Pareto Frontier on GPQA-Diamond

Reasoning accuracy versus token pricing across tiered model dispatch.

Detailed GPQA-Diamond evaluation table →
ModelAccuracy (%)Correct / TotalCost (per 1M)
Enso Ultra98.0194 / 198$25
Enso96.0190 / 198$20
Enso Flash92.9184 / 198$4
GPT-5.6 Sol92.9$25
Dynamic Cascade Dispatch

Speculative verification on high-difficulty reasoning benchmarks. The Flash routing tier achieves 92.9% accuracy at 84% lower token expenditure ($4/M vs $25/M) through selective fallback.

ROUTING TIERS: ULTRA · STANDARD · FLASH
02   RUNTIME AND CONCURRENCY

Process Concurrency & Memory Residency

Heap residency, serialization throughput, and resumption latency under concurrent agent execution.

03   STARTUP AND ISOLATION

Execution Sandbox Cold-Start Latency

Initialization latency across V8 contexts, V8 isolates, and microVM instances without prewarmed standby pools.

Detailed sandbox isolation benchmark →
Process and Sandbox Spawn Latency
Live agent (goroutine)
125 ns
V8 context
0.13 ms
V8 isolate
0.35 ms
MicroVM boot
309 ms
04   MEMORY AND RETRIEVAL

Multi-Hop Retrieval & Fact Reconciliation

Recall accuracy and temporal conflict resolution over extended context windows (up to 262K tokens).

05   CODE RETRIEVAL

AST-Aware Cross-File Dependency Retrieval

Syntactic and dependency graph traversal for code completion on RepoBench-R.

Detailed RepoBench-R code retrieval table →
RepoBench-R →
Cross-file context retrieval for code completion.
18.8 → 33.2
R@1  ·  +14.4 pts
18.8
33.2
Dependency Graph Traversal vs Lexical Baseline

Lexical BM25 retrieval degrades across multi-file repositories due to symbol shadowing. Semantic graph indexing walks import hierarchies and call trees, boosting R@1 from 18.8 to 33.2 (+14.4 points over baseline).

HARNESS: REPOBENCH-R (PYTHON) · R@1 METRIC
06   INFERENCE THROUGHPUT

Prefill & Decode Throughput vs llama.cpp

Throughput evaluation of Hanzo Engine against llama.cpp on identical hardware, weights, and quantization levels.

Detailed cross-backend inference matrix →
BackendModelContextSpeedup
CUDAZen-1.7Bpp40961.088×
CUDAZen-4Bpp20481.051×
Measured under identical FP16/Q4_K quantization on NVIDIA Hopper and Ada architectures.
Fused Attention & Paged KV-Cache

FlashAttention-3 integration and custom fused GEMM kernels prevent memory bandwidth saturation during prefill. Paged KV-cache allocation eliminates memory fragmentation under concurrent batch requests.

BACKENDS: CUDA (SM_90/SM_89) · METAL (APPLE SILICON) · AVX-512
07   REPRODUCIBILITY & PROTOCOL

Evaluation Protocol & Reproduction

All benchmark harnesses, raw run telemetry, and analysis scripts are published in the open repository. Every published figure is directly regenerable from committed test fixtures.

Command Line Reproduction

Clone the evaluation harness and execute the test runner to regenerate raw telemetry and comparison tables from scratch:

Terminal:git clone https://github.com/hanzoai/cloud && cd cloud/bench/brain && make reproduce
Deterministic Execution

Invariant greedy decoding (temperature 0.0), fixed pseudo-random seeds, and uniform prompt templates across all evaluation arms.

Bootstrap Confidence Intervals

95% bootstrap confidence intervals computed over questions and test splits to distinguish empirical signal from stochastic variance.

Hardware Isolation

Bare-metal benchmarking with pinned CPU affinity, isolated memory regions, and dedicated GPU/accelerator contexts.