One model provider. Every machine. The best local configuration the hardware can safely sustain.
Turbofit is a first-class Hermes Agent provider and adaptive local-inference runtime. It inventories physical compute and total usable memory, recommends an evidence-backed ladder, launches a native backend, and exposes one OpenAI-compatible endpoint.
provider: custom:turbofit
model: auto
TurboFit Check scans the machine — dedicated VRAM, unified/integrated memory, or RAM-only — and applies Auto or a selected compatible lane.
Wiki: Full Turbofit Wiki — Model Zoo, benchmarks, hardware tiers, and more.
5 models suitable for running Hermes Agent locally. Ordered by capability.
| Order | Model | Type | Parameters | Optimal Quant | Size | AIME26 | GPQA-D | DeepSWE | LCBv6 | TB4.0 |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Maple Preview | Ternary MoE | 20B-A1B | TQ2_0-head-Q4_K | 5.5 GiB | 87.5% | 73.5% | — | 75.1% | — |
| 2 | Cyber Tiel Coder | Code LLM (MoE) | 35B-A3B | UD-Q4_K_XL | 22.4 GiB | — | — | — | — | — |
| 3 | Qwen 3.8 GSQ-RCO | LLM (Quantized) | 27B | IQ3_S | 11.8 GiB | — | 89.2% | 42.2% | 90.3% | — |
| 4 | Qwen 3.8 Flash Next | MoE (Qwen4 Preview) | 180B / 6B | IQ3_XXS | 47 GiB | — | 91.7% | 58.7% | 91.9% | — |
| 5 | DeepSeek V4.1 Flash | MoE | 763B / 16B | FP8 / UD-Q4_K_XL | 160 / ~40 GiB | — | 90.9% | 74.2% | — | 31.2% |
Fit rule: model_size + KV cache ≤ tier size. KV cache ≈ 2-4GB at 262K context. GSQ-RCO IQ3_S is near-lossless and the default for all GPU tiers.
| Model | Type | System | Parameters | AIME26 | GPQA-D | DeepSWE | LCBv6 | TB4.0 |
|---|---|---|---|---|---|---|---|---|
| Maple Preview ★ | Ternary MoE | Omarchy | 20B-A1B | 87.5% | 73.5% | — | 75.1% | — |
| Cyber Tiel Coder ★ | Code LLM (MoE) | Omarchy, MacBook | 35B-A3B | — | — | — | — | — |
| Qwen 3.8 GSQ-RCO ★ | LLM (Quantized) | Omarchy | 27B | — | 89.2% | 42.2% | 90.3% | — |
| Qwen 3.8 Flash Next ★ | MoE (Qwen4 Preview) | Omarchy | 180B / 6B | — | 91.7% | 58.7% | 91.9% | — |
| DeepSeek V4.1 Flash ★ | MoE | Omarchy | 763B / 16B | — | 90.9% | 74.2% | — | 31.2% |
| Ace Step 1.5 | Music Generation | Omarchy | 4B | — | — | — | — | — |
| Bonsai 27B | LLM (Ternary) | Omarchy | 27B | — | — | — | — | — |
| Cosmos 37B | LLM | Omarchy | 37B | — | — | — | — | — |
| NVIDIA Nemotron 3.5 | Enterprise LLM (MoE) | Omarchy | 30B-A3B | — | — | — | — | — |
| Qwen 2.5 Omni | Multimodal VLM | Omarchy | 3B | — | — | — | — | — |
| Soprano | Creative Micro | Omarchy | 80M | — | — | — | — | — |
★ = on TurboFit List. For full model research, see the wiki.
| Tier | Model | Quant | VRAM Used | Context | Notes |
|---|---|---|---|---|---|
| <8 GB | Maple | TQ2_0 | ~6 GiB | 131K native | SSD streaming |
| 16–64 GB | GSQ-RCO | IQ3_S | ~12 GiB | 262K+ | Near-lossless, low power |
| 96 GB | Flash Next | IQ3_XXS | ~48 GiB | 262K native | Qwen4 preview |
| 128 GB | Flash Next | IQ3_XXS | ~48 GiB | 1M | Max context |
| 256 GB | DeepSeek V4.1 Flash | FP8 | ~160 GiB | 1M | Top capability |
| Tier | Model | Quant | RAM Used | Context | Notes |
|---|---|---|---|---|---|
| <8 GB | Maple | TQ2_0 | ~6 GiB | 131K native | Minimum viable |
| 16 GB | Cyber-Tiel Coder | Q3_K_M | ~14 GiB | 262K native | MoE, fast |
| 32–48 GB | Cyber-Tiel Coder | UD-Q4_K_XL | ~22 GiB | 262K+ | Fast MoE |
| 64 GB | GSQ-RCO | IQ3_S | ~12 GiB | 262K+ | Near-lossless |
| 96+ GB | Flash Next | IQ3_XXS | ~48 GiB | 1M | Top tier |
| Tier | Model | Quant | RAM Used | Context | Notes |
|---|---|---|---|---|---|
| <8 GB | Maple | TQ2_0 | ~6 GiB | 131K native | 218 tok/s on M4 |
| 16 GB | Cyber-Tiel Coder | Q3_K_M | ~14 GiB | 262K native | MoE, fast |
| 32–48 GB | Cyber-Tiel Coder | UD-Q4_K_XL | ~22 GiB | 262K+ | Fast MoE |
| 64 GB | GSQ-RCO | IQ3_S | ~12 GiB | 262K+ | Near-lossless |
| 96+ GB | Flash Next | IQ3_XXS | ~48 GiB | 1M | Top tier |
| Repo | Description |
|---|---|
| SouthpawIN/TurboFit | Main Turbofit runtime |
| deepgrove-ai/mlx-lm-deepgrove | Maple MLX runtime |
| RasoulNik/ssdmoe | SSD MoE streaming |
| tayoun/flash-moe | MoE streaming from SSD |
| NousResearch/hermes-agent | Hermes Agent runtime |
- CPU and Unified Memory prefer MoE over dense — Cyber-Tiel Coder for <48GB, GSQ-RCO at 64GB+
- IQ3_S is the default — near-lossless, lower power, faster inference
- Only 5 TurboFit candidates — Maple, Cyber-Tiel, GSQ-RCO, Flash Next, DeepSeek V4.1 Flash
- Math must be correct — model size + KV cache ≤ tier size
MIT