$ a64forge optimize research-ops --target arm64
Compile the agent for the CPU beneath it.
A64Forge benchmarks every AI-agent stage against local GGUF models, protects its quality floor, and emits a deployable Arm64 routing plan with verifiable evidence.
VERIFIED ARM64 RUNrun-1488b44fa409aarch64
COMPILER OUTPUTDEPLOYABLE
baselineQwen3-4B / Q4_K_Mall stages
optimizedstage-aware routing5 selected routes
- Target
- Linux / aarch64
- Quality gate
- passed
- Source commit
- b414b9e8d52e
WORKFLOW-AVERAGE RESULT
One workload. Same run. Better route.
THE COMPILED WORKFLOW
Five tasks should not inherit one model configuration.
The optimizer evaluated candidates under task-specific quality gates, then selected a Pareto-efficient eligible route for each stage using the balanced preset.
01
classifyQwen3-4B
- quant
- Q4_K_M
- config
- t1 · b64 · c512
- quality
- 1.000
02
extractQwen3-1.7B
- quant
- Q4_K_M
- config
- t4 · b128 · c1024
- quality
- 0.833
03
tool selectQwen3-4B
- quant
- Q4_K_M
- config
- t1 · b128 · c512
- quality
- 1.000
04
reasonQwen3-0.6B
- quant
- Q4_0
- config
- t1 · b64 · c2048
- quality
- 0.833
05
summarizeQwen3-0.6B
- quant
- Q4_0
- config
- t1 · b128 · c2048
- quality
- 1.000
NATIVE HOST EVIDENCE
Measured on Arm. Identified from the machine—not a provider label.
The evidence bundle records the actual architecture, CPU features, runtime tools, configuration, output, and artifact hash.
Read raw results.json →- Architecture
- aarch64
- Logical cores
- 4
- Memory
- 15.57 GB
- CPU features
- NEON · SVE · MATMUL_INT8
- Runtime
- llama.cpp · llama-server
- Performix
- Unavailable — recorded honestly
AUDIT TRAIL
Trace every public claim back to source.
01Verified HTML reportSame run identifier and measured rows↗02Native Arm64 benchmarkGitHub Actions run 31646296185↗03Green validation runTests, build, and evidence checks↗04Evidence manifestRun identity, source commit, and deployment state↗05Compiled routing.yamlDeployable stage-to-model output↗06Apache-2.0 sourceImplementation, tests, and documentation↗
A permanent, read-only publication of a real completed Arm64 benchmark. It does not pretend that a continuously running inference server exists. Re-run the open-source workflow on any compatible Arm64 host to generate fresh evidence.