stories15M_MOE GGUF size and VRAM requirements
ggml-org/stories15M_MOE is a compact language model with 0.04 billion parameters, built on the llama architecture. It is released under the mit license and has been downloaded 82,767 times.
stories15M_MOE is a Mixture-of-Experts model with 4 experts, of which 2 are active on each token. Routing only 2 of 4 experts makes it noticeably faster than a dense model of the same size — but every expert still has to be held in memory, so the numbers below are set by the full parameter count, not the active one. It is built from 6 transformer layers, a hidden size of 288, 6 attention heads.
To run ggml-org/stories15M_MOE locally at a 4,096-token context, its quantized versions need between 0.84 GB (GGUF, lowest quality) and 0.89 GB (F16, highest quality) of memory, weights plus KV cache and a system margin included.
For most users the best balance is F16, needing about 0.89 GB. That means ggml-org/stories15M_MOE fits entirely in the VRAM of a 6 GB GPU or larger, running fully on the GPU.
Available GGUF quantizations for ggml-org/stories15M_MOE include GGUF, Q8_0, F16. The model supports a native context length of up to 256 tokens; a longer context grows the KV cache and the memory needed.
GGUF file size and memory by quantization
Compare real GGUF weight sizes, estimated KV cache and total memory for Q4, Q5, Q8 and every quantization published in this repository.
| Quant. | Bits | Quality | Weights | KV | Total | Speed~ | Verdict |
|---|---|---|---|---|---|---|---|
| GGUF | 3.6 | Fair | 0.02 GB | 0.03 GB | 0.84 GB | 26245.0 t/s | Fits in VRAM |
| Q8_0 | 8.67 | Excellent | 0.04 GB | 0.03 GB | 0.86 GB | 10903.6 t/s | Fits in VRAM |
| F16 | 16.16 | Excellent | 0.07 GB | 0.03 GB | 0.89 GB | 5846.2 t/s | Fits in VRAM |
KV cache computed from the model's exact architecture. Speed is a rough estimate bounded by memory bandwidth.
Frequently asked questions
What kind of model is ggml-org/stories15M_MOE?
ggml-org/stories15M_MOE is a language model with 0.04 billion parameters, based on the llama architecture. It is released under the mit license and distributed as GGUF files for local inference.
Is ggml-org/stories15M_MOE a Mixture-of-Experts (MoE) model?
Yes. ggml-org/stories15M_MOE is a Mixture-of-Experts model with 4 experts, of which 2 are activated per token. That makes it faster than a dense model of the same size, but all 4 experts must be loaded into memory, so the VRAM/RAM it needs is driven by the total parameter count, not the active one.
How much VRAM do you need to run ggml-org/stories15M_MOE?
You need about 0.89 GB of VRAM to run ggml-org/stories15M_MOE entirely on the GPU using the F16 quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.
Can I run ggml-org/stories15M_MOE on an 8 GB GPU?
Yes. With 8 GB of VRAM you can run ggml-org/stories15M_MOE fully on the GPU using F16 (about 0.89 GB).
Can I run ggml-org/stories15M_MOE on a 16 GB GPU?
Yes. With 16 GB of VRAM you can run ggml-org/stories15M_MOE fully on the GPU using F16 (about 0.89 GB).
Can I run ggml-org/stories15M_MOE on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run ggml-org/stories15M_MOE fully on the GPU using F16 (about 0.89 GB).
What context length does ggml-org/stories15M_MOE support?
ggml-org/stories15M_MOE supports a native context length of up to 256 tokens. A longer context grows the KV cache, so it increases the memory needed to run the model.
What is the best quantization for ggml-org/stories15M_MOE?
For ggml-org/stories15M_MOE, higher bits-per-weight means better quality. A common sweet spot is a Q4_K_M or Q5_K_M quantization, which keeps most of the quality while roughly halving the memory versus 8-bit. Pick the highest quantization that still fits in your VRAM.