Solar Pro 2 was a 31-billion-parameter model announced in 2025. Readers evaluating large language models and artificial intelligence should distinguish the May 20, 2025 Solar Pro 2 Preview announcement by Upstage from the older Hugging Face checkpoint used in this article’s original installation instructions. Those instructions did not load Solar Pro 2.
Two 2025 releases, one older checkpoint
Upstage announced Solar Pro 2 Preview on May 20, 2025 and the full Solar Pro 2 launch on July 10, 2025. Its preview announcement described 31B parameters, chat and reasoning modes, a 64K-token context window, and strengths in English, Japanese and Korean. These are dated provider statements; this article does not report a Kingy benchmark run or a current deployment test.
The preview offered a reasoning_effort flag to select structured reasoning rather than ordinary chat. Upstage said its displayed reasoning comparison used general chat benchmarks, rather than the usual math, coding and multi-hop evaluations. That qualification matters when interpreting a claim of parity with another reasoning model.
Upstage reported token savings of 2–30% in certain scenarios, particularly document-heavy or Korean-language tasks. Fewer tokens in those cases do not establish an equivalent reduction in every workload’s latency, total bill or hardware requirement.
Why the original installation instructions were wrong
A model hosted on Hugging Face must be identified by its exact repository and version. The checkpoint linked here is Solar Pro Preview, an older release; the removed installation instructions combined that checkpoint with Solar Pro 2 specifications.
The exact Solar Pro Preview model card identifies a 22B model, not the 31B Solar Pro 2. Its card lists a 4K context limit, English-focused preview coverage, an MIT license and an intended GPU with 80GB of VRAM. Its benchmark table gives a September 8, 2024 release date. None of those properties can be transferred to Solar Pro 2 simply because the names are similar.
API access: do not reuse an unverified model name
The 2025 Solar Pro 2 announcements pointed readers toward Upstage’s managed console. The old curl example requested model solar-pro; the retained announcements do not establish that identifier as a verified Solar Pro 2 endpoint. The example has been withdrawn rather than presented as a working integration.
The link previously labelled as an API-documentation destination is an older Solar Pro product page. Its redirect leads to an older Solar Pro Preview announcement, not a verified current Solar Pro 2 API reference. Check the provider’s current model identifier, access conditions and pricing before integrating a service.
Enterprise deployment was a launch plan
The May preview announcement described cloud-marketplace and on-premises deployment as enhancements planned for the full release. The July launch continued to describe enterprise deployment as an area for further development. These statements do not verify the original claim that this specific model was already available through AWS Marketplace, or that a single-GPU configuration would meet a production workload.
Keep the linked resources attached to the right model
The linked Hugging Face checkpoint documents the older 22B Solar Pro Preview. It is retained as historical source evidence, not a Solar Pro 2 download instruction.
The older Upstage Solar Pro page is also a historical reference. A live page or redirect alone does not establish today’s pricing, availability or an API model alias.
The Upstage GitHub page is a general organization link. This article does not identify a verified Solar Pro 2 repository or claim that its deployment code has been installed or executed.
The earlier blanket statement that Solar Pro 2 was freely downloadable under MIT has been withdrawn. The personally reviewed MIT model card belongs to the older Solar Pro Preview checkpoint. The two Solar Pro 2 announcements do not establish that same license or free downloadable weights for Solar Pro 2.
Read benchmark claims with their conditions
The July 2025 launch described Korean-language benchmarks, including Ko-MMLU, Hae-Rae and Ko-IFEval, and reasoning evaluations such as MMLU-Pro, Math500, AIME and SWE-Bench Agentless. These are provider-reported comparisons tied to particular model configurations, datasets and evaluation methods.
The original Massive Multitask Language Understanding reference identifies a benchmark, not proof of this model’s score. A smaller parameter count is useful context, but does not establish that Solar Pro 2 consistently beats Llama, Qwen or Mixtral across tasks. The unsupported HellaSwag and TruthfulQA superiority claims have been withdrawn.
The preview announcement specifically highlighted English, Japanese and Korean. The July release emphasized Korean performance. This evidence does not establish the original assertion of tested strength across Chinese and multiple European languages, or a general guarantee of accuracy in legal, medical or financial work.
A 64K context allowance in the preview announcement is a capacity claim. It does not demonstrate that every long document or sustained conversation will remain accurate or coherent. A team considering deployment still needs task-level evaluation.
The January 2024 Solar 10.7B review concerns another model. It cannot confirm Solar Pro 2’s performance. The link remains available as historical background, with that boundary made explicit.
Historical availability and current purchase decisions
Solar Pro 2 Preview was offered free on the Upstage Console until July 15, 2025, according to its May announcement. That expired offer is not a current free-access entitlement. The July launch announcement said Solar Pro 2 was available, but this article has not verified today’s account access, price or hosting arrangements.
An older checkpoint’s MIT terms and hardware target do not establish Solar Pro 2’s licensing, redistribution rights, operating cost or commercial access conditions. Review the exact current product terms before relying on any of these for a purchase or deployment.
What the four retained charts show
The images retained from this article’s 2025 publication show an Artificial Analysis-labelled intelligence index, individual evaluation results, a cost comparison and output-token counts. They are historical image evidence, not a fresh leaderboard check or a benchmark reproduced by Kingy.
The first chart includes a legend for estimated values awaiting independent evaluation. Its Solar Pro 2 bar is solid purple; the patterned estimate marking does not appear on that bar. The detailed evaluation image also contains a legend for results claimed by an AI lab. Those printed qualifications should be preserved without assigning them to a particular model’s result where the marking does not support that reading. Neither image establishes today’s ranking or a fresh Kingy evaluation.
The cost image compares the cost of running that intelligence-index evaluation. It does not quote a current API subscription, token tariff or a representative production bill. The token-count image measures the displayed evaluation workload; it does not establish a universal per-query cost advantage.
Practical evaluation questions
For a historical release comparison, identify the exact model and chat or reasoning configuration first. Then compare the same evaluation protocol, language, tools and inference budget. A vendor demonstration or one aggregate chart does not decide which model is best for a particular organization.
Customer support, document analysis, translation and tool-assisted workflows are plausible evaluation areas. The original claims that unnamed customers had already achieved lower costs and faster responses are not supported by the reviewed announcements and have been withdrawn.
This article offers no hands-on installation, throughput measurement, security audit or production incident record. Its corrected account separates the dated provider announcements from the older checkpoint and the retained chart claims.
Before paying for or deploying a model
Check the exact model/version, license, context and output limits, data-handling terms, region availability and current pricing. The older Solar Pro product page does not establish those details for a current Solar Pro 2 service. This article has not submitted API requests or committed funds to test access.
A useful release, with limits on the evidence
The 2025 announcements presented a compact model with multilingual support and a reasoning mode. That is the supported historical story. It does not require the original claims of universal benchmark superiority, guaranteed single-GPU savings or verified enterprise outcomes.
Sources and model identity
The July 10, 2025 Solar Pro 2 launch announcement documents the full release. The older Hugging Face model card documents the separate 22B Solar Pro Preview checkpoint. Keeping those identities distinct prevents an installation guide from promising capabilities its code never loads.
The official Solar Pro 2 launch page and older Solar Pro Preview model card remain linked for inspection. The original publication date and URL are preserved. This correction does not certify a current purchase option or report any model execution.
The Kingy Brief
Get The Kingy Brief.
AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.
Free · Double opt-in · Unsubscribe anytime
Sign up for The Kingy Brief on Beehiiv.
Signup help and newsletter schedule
Signup form provided by Beehiiv. After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.
Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.