Identity
Model family, parameter class, architecture, interface, and exact weight revision name the object under evaluation.
WhitepaperREC / 01
A public method for model claims.
This paper defines the record Ainfera uses to describe a Neptune release and interpret its evaluation. It binds model identity, workload, runtime conditions, observation, and uncertainty into one inspectable object.
Read the record01Record
Four fields before one number.
A benchmark is useful only when another reader can reconstruct what was measured and why the comparison is valid. The record therefore exists before the result.
Workload
The task boundary, suite, data version, scoring rule, peer set, and exclusions define the work being measured.
Runtime
Hardware, memory, serving format, context, quantization, and sampling settings describe the operating conditions.
Observation
The result carries its date, repetitions, uncertainty, exceptions, and source record. Unobserved values remain N/A.
02Reference object
Neptune 27B, in context.
Neptune 27B is an open dense model in the approximately 27-billion-parameter class. It is intended as the reasoning core of private agent systems, where tool use, sustained state, long-horizon execution, and recovery form one operating problem.
System categories describe intended use. They are not measured performance claims.
- Parameter class
- 27B
- Architecture
- Open dense
- Interface
- Agent-native
- Deployment intent
- Private systems
- Public results
- N/A
03Publication state
The table exists before the score.
Public benchmark data for Neptune 27B is not yet available. When results are released, each value will carry the exact model, suite, runtime, date, and uncertainty needed to interpret it.
Benchmark
No number without its conditions.
Public benchmark data for Neptune 27B is N/A. When results are available, every value will carry the exact model, suite, runtime, date, and uncertainty needed to interpret it.
Evaluation register / Neptune 27BRelease field · values publish only with complete conditions
Public stateN/A
01 / AFRAgent finish rateComplete task episodes
02 / TJVTool JSON validitySchema-valid calls and arguments
03 / RAFRecovery after failureRepair after rejected or malformed calls
04 / THRThrash ratioRepeated or unnecessary action loops
05 / VACVAC per cost and timeValue-adjusted completion by envelope
01ModelExact revision
02SuiteNamed workload
03RuntimeHost and format
04DateObserved window
05ResultObserved or N/A
Method / Rev. 01Evaluation philosophy