DEV.co
Open-Source LLM · upstage

solar-pro-preview-instruct

Solar Pro Preview is a 22-billion-parameter instruction-tuned language model from Upstage designed to run on a single 80GB GPU. It is a pre-release version with English-only support and a 4K context window. According to published benchmarks, it performs comparably to much larger models (e.g., Llama 3.1 70B) on several tasks, particularly instruction-following and mathematical reasoning. The model uses MIT license and is openly available without gating.

Source: HuggingFace — huggingface.co/upstage/solar-pro-preview-instruct
22.1B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
44.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerupstage
Parameters22.1B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads44.9k
Likes458
Last updated2024-09-20
Sourceupstage/solar-pro-preview-instruct

What solar-pro-preview-instruct is

22B parameter dense transformer trained via depth up-scaling of Phi-3-medium (14B). Instruction-tuned using ChatML template. Supports text generation via standard transformers library with trust_remote_code=True. Estimated 80GB VRAM requirement for float32 inference; quantization strategies not documented. Pre-release status with limited language coverage and 4K context length; full release planned for November 2024 with expanded language support and longer context.

Quickstart

Run solar-pro-preview-instruct locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="upstage/solar-pro-preview-instruct")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.

Deployment

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

Single-GPU Inference Deployment

Fits within a single 80GB A100 or H100 for inference without model sharding, reducing operational complexity and latency for inference-heavy workloads.

Instruction-Following and Q&A Systems

Tuned for conversational and instruction-following tasks. Benchmark data shows strong IFEval (84.37) and GSM8K (89.69) scores, making it suitable for customer support, tutoring, and guided reasoning applications.

English-Only Enterprise Deployments

MIT license and ungated model are clear choices for organizations needing straightforward, transparent licensing for internal or commercial LLM-based applications without legal friction.

Running & fine-tuning it

Estimated 80GB VRAM for float32 inference (as stated in model card). Quantization (int8, fp16, AWQ) would reduce footprint but is not documented. Inference framework must support trust_remote_code=True (transformers 4.44.2+ stated in example code). Training or fine-tuning not documented; likely requires A100 80GB or H100 for reasonable throughput.

Not addressed in model card. LoRA/QLoRA feasibility unknown; 22B model on single 80GB GPU leaves <10GB margin for training. Standard SFT fine-tuning would likely require 2× 80GB+ GPUs or multi-GPU setup. Consider quantization (4-bit QLoRA) to fit on single GPU, but custom_code flag indicates non-standard architecture—test thoroughly before committing to fine-tuning pipeline.

When to avoid it — and what to weigh

  • Multi-Language Requirements — Model is English-only in pre-release form. Official version planned for November 2024. Not suitable for production multi-lingual systems today.
  • Long Context / Document Processing — Maximum context length is 4K tokens. Unsuitable for processing long documents, extended conversations, or RAG systems requiring >4K token windows.
  • Production Systems Without Stability Validation — Pre-release label and 2024-09-20 date indicate early status. Organizations requiring stable, mature LLM deployments should validate extensively or wait for official November 2024 release.
  • Fine-Tuning with Limited Hardware — 22B model on 80GB GPU leaves minimal headroom for training-time optimizations (LoRA, QLoRA). Full fine-tuning requires additional memory or multi-GPU setups.

License & commercial use

MIT license. Permissive OSI-approved license allowing free use, modification, and redistribution for any purpose (commercial, proprietary, closed-source) with attribution.

MIT license explicitly permits commercial use without restriction. No gating, no API key requirement for downloads, no commercial terms embedded. Deploy in production without legal review beyond standard MIT compliance (retain license header). No enterprise agreement or SLA documented; direct support likely through HuggingFace discussions or Upstage channels.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceModerate
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

Model includes trust_remote_code=True flag, indicating custom Python execution during loading. Review custom code before deploying in restricted environments. No adversarial robustness, jailbreak resistance, or safety-layer documentation provided. Pre-release status may imply less rigorous security vetting. Standard LLM risks apply: prompt injection, data leakage in training, and model inversion attacks are not addressed. Assume responsibility for prompt validation and output filtering in production.

Alternatives to consider

Phi-3.5-MoE

41.9B MoE (6.6B active), MIT license, similar release timeline. Larger effective capacity but higher compute cost and more complex routing. Consider if task complexity warrants additional parameters.

Llama 3.1 8B

8B dense model, Llama license (requires review for commercial use), can fit on consumer GPUs (40GB). Smaller, simpler, but lower benchmark performance. Good for cost-sensitive or edge deployments.

Gemma 2 27B

27B dense model, Gemma license (requires review for commercial use), strong IFEval (75.36). Larger than Solar Pro, requires 80GB+ GPU, but established documentation. Consider if single-GPU constraint is flexible.

Software development agency

Ship solar-pro-preview-instruct with senior software developers

Our AI engineering team can help you architect single-GPU inference pipelines, fine-tuning strategies, and production monitoring for Solar Pro. Contact us to discuss your LLM deployment requirements and explore whether Solar Pro or an alternative fits your use case.

Talk to DEV.co

Related 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.

solar-pro-preview-instruct FAQ

Can I use this model commercially in production?
Yes. MIT license permits unrestricted commercial use, including deployment, modification, and resale without royalties or legal friction. No commercial agreement or API key registration required; download and deploy as-is. Ensure proper attribution per MIT terms.
What GPU do I need to run Solar Pro Preview?
A single 80GB GPU (A100 or H100) is the stated requirement for full-precision (float32) inference. Quantization (fp16, int8, 4-bit) may reduce footprint but is not documented. For fine-tuning, expect to need 2× 80GB GPUs or use multi-GPU distributed setup.
What is the maximum context window?
4K tokens in this pre-release. Official Solar Pro (November 2024) will support longer context. If your application needs >4K, use alternative models or wait for the official release.
Does the model support system prompts?
No. Current version uses ChatML template for user and assistant roles only. System prompts will be added in the official November 2024 release. Work around by embedding system instructions in the first user message if needed.

Custom software development services

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If solar-pro-preview-instruct is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Solar Pro Preview?

Our AI engineering team can help you architect single-GPU inference pipelines, fine-tuning strategies, and production monitoring for Solar Pro. Contact us to discuss your LLM deployment requirements and explore whether Solar Pro or an alternative fits your use case.