NEPTUNE-1.0-27B-AGENT IS IN TRAINING — FOLLOW THE RUN ON HUGGING FACE  →
AinferaainferaHugging FaceFOLLOW ON HF
++
MODELS · STAGE 01 · FLAGSHIP

Neptune-1.0-27B-Agent

IN TRAINING

The agent-native flagship, on the line now. Hermes-native tool calling, multi-step agentic workflows, and a name that is the spec — 27 billion parameters, true count. Weights publish the moment the gate passes them.

FOLLOW THE TRAINING RUN ↗HOW THE GATE WORKS →
APACHE 2.0 AT RELEASE · 27B TRUE COUNT · HERMES-NATIVE
SFT ✓DPO ✓EVAL GATESTAGE 03/06 · GRPO
TRAINING RUN · STEP 114,208 / 300,000ILLUSTRATIVE
01 · SPEC

The name is the spec

PARAMETERS27B · TRUE COUNT
CONTEXT128K TOKENS
TOOL FORMATHERMES-NATIVE
BASENEPTUNE-1.0 PRETRAIN · DENSE
LICENSEAPACHE 2.0 AT RELEASE
STATUSIN TRAINING · GRPO
02 · TRAINED FOR

Agent work, not chat

HERMES-NATIVE TOOL CALLINGStructured tool calls in the format agents already speak — no wrapper prompts, no parsing glue.
MULTI-STEP WORKFLOWSGRPO over full agent episodes: plans, function chains, and long-horizon task completion.
FAILURE RECOVERYTrained on recovery from failed calls and bad tool output — the part of agent life chat models never see.
BENCHMARKS PUBLISH WITH THE EVAL CERTIFICATE AT RELEASE — TARGETS ONLY UNTIL THE GATE SIGNS THEM.