Every model from the bench, against the reference rigs: the one quant we would actually download for each tier, its file size, and a plain speed class. Fits are at 32k context, from file size plus runtime overhead (our estimate) — computed, not measured. When a tier doesn’t fit, the matrix says so.
Hybrid MoE: 24.0B total parameters, 2.45B active per token, 131,072-token context, text only. Tuned for agents and tool calling.
Released October 13, 2026. License: LFM Open License v1.0 — commercial use yes under $10M annual revenue; paid license required above.
Model card · Official GGUF builds · License
Runs in llama.cpp, MLX, vLLM, LM Studio; official GGUF and MLX builds on Hugging Face. Use a current llama.cpp build (b11606+) and pass --jinja so the embedded chat template drives tool calls. Vendor sampler settings ship in the file: temperature 0.1, top_k 50, repetition penalty 1.05. Difficulty: light setup.
| Reference rig | Recommended quant | File size | Why this one | Speed class |
|---|---|---|---|---|
| 8GB GPU | No full-GPU fit | — | CPU expert offload only. | Usable |
| 16GB GPU | IQ4_XS | 13.30GB · m9ter1a | Near-BF16 tool calls in community tests. | Comfortable |
| 24GB GPU (4090 class) | Q5_K_M | 17.37GB · m9ter1a | Closest to BF16 of the measured files. | Flies |
| 48GB GPU | Q8_0 | 25.49GB | Near-lossless, room for long context. | Flies |
| Mac 32GB | Q4_K_M | 14.53GB · official | Safe pick, unified-memory headroom. | Comfortable |
| Mac 64GB | Q8_0 | 25.49GB | Near-lossless; 184 tok/s on M5 Max at Q4_K_M (vendor). | Flies |
| 96GB class | BF16 | 47.96GB | Full precision, full context. | Flies |
| 128GB unified | BF16 | 47.96GB | Full precision, full context. | Flies |
| CPU offload | Q4_K_M | 14.53GB | Expert offload on — 2.45B active keeps it usable. | Usable |
Speed classes are the bench column’s own words: Flies, Comfortable, Usable. Fits are our estimate at 32k context.
Steal this prompt — the fill-in version
I want to run [MODEL NAME] locally on my machine. My machine: [GPU OR MAC MODEL], [VRAM OR UNIFIED MEMORY, e.g. 24GB], [SYSTEM RAM], [OPERATING SYSTEM]. The recommended quant for my memory tier is [QUANT, e.g. Q5_K_M]. Setup notes for this model: [PASTE THE SETUP NOTES — runtime, required flags, sampler settings]. Give me exact, step-by-step install and run commands for my machine, sized for a context length of [CONTEXT, e.g. 32k]. If a step depends on something you do not know about my machine, ask me first — and flag anything you have to guess rather than inventing a step, a file name, or a flag.
Copy it, fill the brackets with your own machine, and paste it into a frontier model. The last lines are the important part: the model is told to ask about anything it doesn’t know and to flag its guesses instead of inventing steps.