SOLAR 10.7B v1.0
SOLAR 10.7B is a base pretrained model from Upstage built by applying depth up-scaling (DUS) to Mistral 7B, pushing parameters to 10.7B without a traditional MoE architecture. It scored competitively on the H6 benchmark against models up to 30B params, including Mixtral 8x7B. This is a raw base model — it will not follow instructions or chat without additional fine-tuning.
If you need a compact, commercially licensable base model to fine-tune on your own data, SOLAR 10.7B is a legitimate starting point — the DUS architecture is real and the benchmark results hold up. For Korean-language tasks specifically, treat it as an unknown quantity until you run your own evals. Skip it if you need a ready-to-run chat model; nothing works out of the box here. Hedge: worth pulling if you have a fine-tuning workflow ready, but not a casual download.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.05/10. License is explicitly apache-2.0 in the card and correctly flagged commercial-ok. Params (10.7B), family (llama arch), vendor (upstage), and base-model status all match the card. The description honestly flags it as a base model requiring fine-tuning, and the verdict is appropriately hedged. The 'korean' useCase tag is weak — the card doesn't claim Korean specialization for v1.0, and the row itself acknowledges Korean capability is uncharacterized, so including 'korean' in useCases is inconsistent. Context length of 4096 is reasonable for a Mistral-derived model but not explicitly stated in the excerpt. Overall a solid, honest row that just clears the bar.
Flags: - useCases includes 'korean' but the row's own weaknesses section says Korean capability is uncharacterized — inconsistent signaling - contextLength 4096 not explicitly confirmed in the provided card excerpt (DeepSeek confidence 0.5)
Overview
SOLAR 10.7B is a base pretrained model from Upstage built by applying depth up-scaling (DUS) to Mistral 7B, pushing parameters to 10.7B without a traditional MoE architecture. It scored competitively on the H6 benchmark against models up to 30B params, including Mixtral 8x7B. This is a raw base model — it will not follow instructions or chat without additional fine-tuning.
Strengths
- Depth up-scaling on Mistral 7B yields strong benchmark results at 10.7B params
- Outperforms some sub-30B models on H6 benchmark, including Mixtral 8x7B on select tasks
- Apache-2.0 license — clean for commercial fine-tuning and deployment
- Reasonable VRAM footprint for a capable base model
Weaknesses
- Base model only — unusable for chat or instructions without fine-tuning
- 4096-token context is short by current standards; long documents will be truncated
- Korean language capability is uncharacterized — do not assume multilingual strength
- HF download count (15K) suggests limited community validation relative to comparable base models
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 5.9 GB | 8 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of SOLAR 10.7B v1.0.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run SOLAR 10.7B v1.0?
Can I use SOLAR 10.7B v1.0 commercially?
What's the context length of SOLAR 10.7B v1.0?
Source: huggingface.co/upstage/SOLAR-10.7B-v1.0
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify SOLAR 10.7B v1.0 runs on your specific hardware before committing money.