Solar Open 100B

What it is

Solar Open 100B is an open-weight large language model from the Korean company Upstage. Its weights were uploaded to Hugging Face on 10 December 2025^[1]. It is a mixture-of-experts (MoE) model with 102.6 billion parameters in total, of which 12 billion are active for each token the model processes^[1]. It works in both English and Korean^[1].

Mixture of experts is a way of making a very large network affordable to run. Rather than sending every token through every part of the network, an MoE model keeps many separate sub-networks — the experts — and a routing step decides which of them each token visits. The total parameter count then describes how much the model holds, while the active count describes what one token costs to compute. Here the gap between the two numbers is about eight and a half to one^[1].

How it is built

The expert layout is specific: 129 experts in total, 128 routed and one shared, with each token sent to the top eight^[1]. The shared expert is the one every token passes through regardless of routing, while the routed ones are chosen per token — a common arrangement that gives the router a fixed baseline to work against.

Pre-training used 19.7 trillion tokens^[1], and the run took place on NVIDIA B200 GPUs^[1]. Upstage says the model was trained entirely from scratch rather than continued from another checkpoint^[1], so the released weights come from a run that began with random initialisation instead of being built on top of an existing model.

The context window is 128K tokens^[1], enough to hold a long document or a substantial code file in a single prompt.

Licence and how to run it

The weights ship under the Upstage Solar License, a custom licence rather than an OSI-approved one^[1]. The difference is worth knowing before downloading: open weights mean the files are downloadable, while an open-source licence is a specific legal category, and a custom licence can attach conditions an OSI licence would not permit. Anyone planning commercial deployment has to read the terms rather than assume.

Two things make the release easier to use in practice. NotaAI published an official INT4 quantised version^[1]. Quantisation stores the weights at lower numerical precision — 4-bit integers instead of the 16-bit format used in training — which cuts memory use at some cost in accuracy. The minimum hardware listed is four A100 80GB GPUs^[1], a figure that reflects the fact that a sparse model still has to hold all 102.6 billion parameters in memory even though only 12 billion are used per token.

Report and successor

The technical report followed on 11 January 2026, about a month after the weights^[2]. It describes synthesising 4.5 trillion tokens for underserved languages^[2] — generating training text for languages that are thinly covered in material collected from the web, instead of relying only on what already existed.

A larger model, Solar-Open2-250B, followed on 22 July 2026^[3].

What to take from it

Three things about this release are worth noting rather than one. First, the language pair: the model is built for English and Korean^[1], not for English alone. Second, the licence: a custom licence means the boundary of "open" has to be judged from the terms, not from the label^[1]. Third, the data work is documented — the report's account of 4.5 trillion synthesised tokens describes how coverage was added for languages that had little of it^[2]. For anyone who wants to try the model, the official quantised build and the stated minimum hardware^[1] make it possible to estimate the cost before starting.