stories15M_MOE
stories15M_MOE is a 36M-parameter mixture-of-experts (MoE) model built by repeating a 15M TinyLLaMA variant four times as separate experts. It is explicitly labeled as a testing/experimental model, not intended for production use. The model generates short children's stories. It supports LoRA fine-tuning and is available in multiple formats (safetensors, GGUF). MIT license, freely available, no gating.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | ggml-org |
| Parameters | 36M |
| Context window | Unknown |
| License | mit — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 65.6k |
| Likes | 7 |
| Last updated | 2024-08-05 |
| Source | ggml-org/stories15M_MOE |
What stories15M_MOE is
A MoE text-generation model with 4 expert branches initialized from ModelCloud/tinyllama-15M-stories, totaling ~36M parameters. Router initialization is random. Context length is not documented. Available in Transformers, safetensors, and GGUF formats. Includes a LoRA adapter trained on Shakespeare text. Compatible with HuggingFace Endpoints and text-generation-inference services.
Run stories15M_MOE locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="ggml-org/stories15M_MOE")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
ESTIMATE (unconfirmed): 36M parameters in FP32 ≈ 144 MB, FP16 ≈ 72 MB, GGUF quantization ≈ 20–40 MB depending on quantization level. Inference on CPU or edge devices (e.g., Raspberry Pi, mobile) likely feasible; verify actual memory footprint with target quantization format.
LoRA fine-tuning is feasible and demonstrated in model card (Shakespeare adapter). QLoRA likely viable on constrained hardware due to small parameter count. Full fine-tuning should require minimal VRAM. No multi-GPU training required. Training code and best practices not documented; recommend consulting HuggingFace fine-tuning guides.
When to avoid it — and what to weigh
- Production or High-Reliability Systems — Model card explicitly states it is not intended for production use. Output quality is inconsistent; router initialization is random and not optimized.
- Complex Reasoning or Knowledge-Heavy Tasks — At 36M parameters with story-generation training, this model lacks the capacity for Q&A, summarization, code generation, or domain-specific reasoning.
- Systems Requiring Known Context Length — Context length is undocumented. Cannot reliably estimate maximum input size or suitability for long-document processing.
- Safety-Critical Applications — No safety alignment or content filtering documented. Unsuitable for customer-facing or regulated use without extensive additional testing and guardrails.
License & commercial use
MIT License. Permissive OSI-compliant open-source license. Allows use, modification, and distribution with proper attribution.
MIT is a permissive license that permits commercial use, including in proprietary products. However, the model card explicitly disclaims production use—intended for testing only. Commercial use should include internal assessment of fitness and risk, especially given the stated non-production status. Recommend documenting rationale if deploying commercially.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Moderate |
| Documentation | Limited |
| License clarity | Clear |
| Deployment complexity | Low |
| DEV.co fit | Possible |
| Assessment confidence | High |
No security analysis, content filtering, or adversarial robustness testing documented. Random router initialization may introduce non-deterministic behavior. No discussion of data leakage, prompt injection mitigations, or alignment. For safety-sensitive applications, conduct independent red-teaming and content policy review before deployment.
Alternatives to consider
TinyLLaMA (4.1B)
Larger, better-trained baseline with production intent. Suitable if story generation or lightweight inference is needed with better quality and documented context length.
Qwen1.5-0.5B or Phi-2
Similarly small but optimized for general-purpose text generation, coding, and reasoning. Better documentation and release maturity.
Mistral-7B (with quantization)
Higher capability, mature open-source option for production scenarios. Larger but widely tested and documented.
Ship stories15M_MOE with senior software developers
This model is ideal for testing MoE architectures and fine-tuning experimentation on edge devices. Not recommended for production. Review the model card, verify context length for your use case, and test LoRA adapters in your environment. If you need a production-grade alternative, consider TinyLLaMA or Qwen1.5.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
stories15M_MOE FAQ
Can I use this model commercially?
What are the hardware requirements?
Is this model suitable for my production chatbot?
Can I fine-tune this model?
Software development & web development with DEV.co
Adopting stories15M_MOE is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.
Evaluate stories15M_MOE for Your Use Case
This model is ideal for testing MoE architectures and fine-tuning experimentation on edge devices. Not recommended for production. Review the model card, verify context length for your use case, and test LoRA adapters in your environment. If you need a production-grade alternative, consider TinyLLaMA or Qwen1.5.