Source: Ollama
Instruction Document AI 10.7B Open Weights
Solar is an open-weight model family from Upstage, a Korean AI company known for its document understanding products. The flagship Solar 10.7B uses a technique called depth upscaling — essentially merging two Mistral 7B models and continuing training — to produce a 10.7B model that punches well above its weight class. When released, it topped the HuggingFace Open LLM Leaderboard for its size tier. Solar Pro is a larger, higher-capability variant for enterprise deployments.
Solar is a practical choice if you want a capable local assistant that runs well on modest hardware. The original Solar 10.7B model offers solid reasoning and clean output for everyday writing, summarization, and Q&A. While it remains efficient, it has been surpassed by newer compact models on most benchmarks. For privacy-focused users who value low resource usage, it is still a sensible option for general tasks.
Teams that need grounded, on-premise AI for document-heavy workflows — such as internal knowledge search, drafting, and structured summarization — will find Solar useful due to its reputation for staying faithful to provided context. The smaller models keep deployment costs predictable, while the newer Solar Open 100B MoE provides significantly higher capacity for organizations that need more performance without sending data to the cloud. A focused pilot is recommended before standardizing on the larger MoE model.
Developers building internal tools, RAG pipelines, or applications that require predictable behavior will appreciate Solar's efficiency and groundedness. The 10.7B size is light enough to embed in apps without large GPU budgets. The Solar Open 100B MoE offers more headroom for demanding workloads while maintaining efficient inference thanks to its mixture-of-experts architecture. Plan for the increased memory and serving requirements if moving beyond the compact sizes.
Solar is interesting for researchers studying model efficiency and architecture trade-offs. Upstage's Depth Up-Scaling approach produced strong results for its size, making the 10.7B model a useful reference point for reproducible experiments on limited hardware. The newer Solar Open 100B MoE provides a valuable data point for studying large open mixture-of-experts behavior, though community tooling and findings are still maturing.
| Model | Parameters | VRAM (4-bit) | Notes |
|---|---|---|---|
| Solar 10.7B Instruct | 10.7B | ~7GB | Flagship open-weight Solar; excellent quality/size ratio |
| Solar Pro Preview | 22B | ~14GB | Higher capability; enterprise document focus |
Solar 10.7B Instruct runs comfortably on a g4dn.xlarge (16GB VRAM). Use Ollama:
ollama run solar:10.7b-instruct-v1-q4_K_M
Q4_K_M GGUF fits easily on a 12GB GPU. For CPU-only machines, the GGUF runs acceptably on 16GB RAM systems via llama.cpp.
Solar Pro (22B) runs well on a single A10G (24GB) or two RTX 4090s. Serve with vLLM for multi-user access.
| Model Size | Ollama Tag | Recommended NoCloudGPT Tier | Notes | |
|---|---|---|---|---|
| Solar 10.7B | solar:10.7b |
Starter / Standard | Efficient and well-rounded for its size. Strong for personal use, RAG, and lightweight team tools, though newer compact models now lead on benchmarks. | Get quote → |
| Solar Open 100B MoE | solar:100b |
Professional / Enterprise | High-capacity mixture-of-experts model with efficient inference relative to its size. Newer to the open ecosystem — a focused pilot is recommended before large-scale deployment. | Get quote → |
Positioning note: The Solar 10.7B remains a capable and efficient model for many general-purpose tasks. The newer Solar Open 100B MoE offers significantly more capacity for demanding workloads. For maximum performance on current benchmarks, newer families (Llama 3.3, Qwen 2.5, DeepSeek V3) are generally stronger.
Licensing for Solar models varies by specific release. The widely used Solar 10.7B has historically been available under relatively permissive terms. The newer Solar Open 100B MoE has its own license that should be reviewed directly, as commercial use rights and redistribution conditions can differ between versions.
Because Upstage maintains multiple Solar releases with distinct terms, always confirm the exact license on the official model card for the specific size you intend to deploy.
Official source:
Artwork and source references
Hero image
Developer reference
Solar is developed by Upstage AI. NoCloudGPT is an independent deployment service not affiliated with Upstage AI.
Model catalog
See Ollama library: solar for current tags and sizes.