Turbofit

What Turbofit does

TurboFit Check scans whether the machine has dedicated VRAM, unified or integrated memory, or RAM only, then applies either an automatic or a manually selected compatible model tier from the TurboFit List, a ranked, benchmark-scored lineup of five models suitable for running Hermes Agent locally, from a 5.5 GiB ternary MoE model up to a 160 GiB DeepSeek V4.1 Flash deployment. The fit rule is explicit: model size plus KV cache must stay within the tier's size budget, with hardware tiers documented separately for GPU VRAM, system RAM and Apple Silicon unified memory.

Configured in Hermes as provider: custom:turbofit with model: auto, Turbofit then picks and launches the matching native backend itself. A fuller model zoo beyond the five-model TurboFit List, along with benchmark scores (AIME26, GPQA-D, DeepSWE, LiveCodeBench v6 and TB4.0) and hardware-tier tables, is documented in the project's wiki.

Key features

• Hardware inventory (dedicated VRAM, unified memory or RAM-only) with an evidence-backed model recommendation
• Five-model TurboFit List ranked by capability, from a 5.5 GiB model up to 160 GiB
• One OpenAI-compatible endpoint exposed regardless of which native backend is launched
• Separate hardware-tier tables for GPU VRAM, system RAM and Apple Silicon unified memory
• Benchmark scores (AIME26, GPQA-D, DeepSWE, LiveCodeBench v6, TB4.0) per model

When to use it

• Letting Hermes Agent pick the best model a given machine can run, without manual tuning
• Running Hermes fully locally on hardware ranging from under 8GB VRAM up to 256GB RAM
• Comparing local model options for Hermes by benchmark score before committing to one

Who it is for: Hermes Agent users who want to run models locally and have the right model for their hardware chosen automatically rather than guessed.

How it fits with Hermes Agent

Turbofit is described as a first-class Hermes Agent provider, configured directly as provider: custom:turbofit inside Hermes' own provider configuration.

FAQ