Changelog
Roadmap

Have something to say?

Tell us how we could make the product more useful to you.

The API service is still in beta testing.

I'm using the Chat Completion API to request structured output, but the response is often truncated without any warning ("finish_reason":"stop"). JSON escaping often fails. Also, whether it’s because I just created the API key or because I’ve hit the rate limit, requests that initially returned a response within 10 seconds now often take over 100 or 200 seconds under the same conditions. I’ve also registered a payment method, but it doesn’t seem to make a difference. I gave up while running benchmarks to debug the "structured output" error. The model’s performance doesn’t seem bad, but the service quality makes it difficult to put it to immediate use. For reference, all service statuses are showing green. I started with `gemini-2.5-flash` to try out a Korean model or one served domestically, but switched to `solar-pro4`—it’s been quite a struggle. The sorrow of a citizen of a developing country ㅠ

노영선 8 days ago

2
⚙️

Upstage Console

Error in logic processing

Hi Upstage team, I wanted to share a reasoning issue I found while benchmarking Solar Pro 4 Thinking. I think it is worth looking at because this was not simply a formatting mistake or a bad final transcription, the model made a genuine logical error in both its corrected code and its follow-up explanation. I would also like to note that Ollama small q4 models (Gemma 4 31b, Qwen 27b) are able to solve this coding problem. I also tried increasing the reasoning to MAX on Open Router and it kept making the same error. The benchmark asks the model to debug a meeting-room scheduler containing exactly seven logical bugs. One of the rules is that every meeting requires exactly 15 minutes of cleanup afterward, meaning the conflict check must work symmetrically whether the new meeting comes before or after an existing booking. The correct condition is: not ( req["end"] + 15 = booking["end"] + 15 ) Solar Pro 4 instead produced: not ( req["end"] = booking["end"] + 15 ) This is asymmetric. It gives the existing booking its required cleanup period, but does not give the new request its cleanup period when the new request occurs first. That directly caused Solar to incorrectly place request R6 into Birch: R6 runs from 585–600. Birch already contains R2 starting at 600. Solar considered this valid because 600 . But R6 requires cleanup until 615, so R2 cannot start at 600. Birch is therefore unavailable, and R6 must instead be placed in Cedar. I later asked Solar why it had put R6 in Birch. Rather than recognizing the mistake, it explicitly defended the result and said: "no conflict with R2 (R2 starts at 600, after R6 ends at 600)." That follow-up is important because it confirms that this was not an accidental output mistake. The model genuinely failed to apply the cleanup constraint symmetrically. Most of the rest of the solution was strong: Solar correctly fixed the request ordering, capacity comparison, equipment comparison, deep-copy issue, room-selection ordering, and chronological output sorting. But this particular failure is significant because the benchmark specifically tests whether the model can maintain program-state constraints and correctly repair logical code. For a new model in this size/capability class, I would consider this more serious than a minor benchmark miss. The rule itself is simple and explicitly stated in the prompt, yet the model: failed to implement it correctly, produced an incorrect downstream program state, and then confidently defended the incorrect reasoning when challenged. That combination, incorrect implementation plus confident persistence after the error is pointed out, is the part I think deserves attention. I am excited to see a new contender like Sol Pro 4 in the medium sized class and will be happy to continue to test your model. I hope this is useful for evaluation and debugging. I can also provide the full benchmark prompt if that would help reproduce the issue.

An Anonymous User 8 days ago

1
💡

Upstage Studio

It would be great if we could enter all the correct answers for the Feedback Loop at once!

I'm currently training a synthetic document using the Feedback Loop feature, but I'm leaving this feedback because I've encountered one inconvenience while using it :) Currently, it seems that a warning triangle appears next to each field with a low confidence score in the extraction results, and you have to click on each one individually to confirm it as correct before it’s added to the correct answer data. However, in practice, warning indicators often appear even on values that are already correct, so as the number of documents to train increases, manually verifying each one is becoming increasingly tedious…! I’m leaving this feedback because I think it would be great to have an option to pre-enter the correct values for each document (by uploading a CSV or JSON file) and apply them all at once when the correct values are already known! That would make the schema improvement process—from the Feedback Loop to QuickTune—much smoother :) Thank you :)

[조구만 루키] 길연진 9 days ago

💡

Upstage Studio

Completed

It looks like the QuickTune result notification and the information in the QuickTune results don't match!

The QuickTune result notification says, “Extraction quality has improved,” but but when I click that notification to view the QuickTune results, it says, “No keys with improved accuracy were found.” Since this seems to happen frequently, it’s hard to tell whether the extraction has actually improved or not. Please look into this! Thank you :)

[조구만 루키] 길연진 9 days ago

2
💡

Upstage Studio

Completed

Is there a way to obtain a Solar Open 2 API key?

Like everyone else, I’ve been using DS V4 for a while now, and I use a variety of models—including Kimi, MIMO, OpenCode, GLM, and more—through my annual subscription. At first, just looking at the benchmark results, I thought, “Does this even mean anything to me?” But then I ran two problems through Kimi K2.6, DS V4 Flash, Solar Open2, and GPT5.6 Sol—and Solar Open2 and GPT5.6 Sol got them exactly right. This made me wonder if their actual performance in coding might be even better than what the benchmark results suggest. I’m also active in the community of a U.S.-based startup that provides OpenWeight APIs, and while discussing this, https://open2-beta.upstage.ai/ , but while some people got a 403 error, others were able to access it. I wonder if there are countries where the page doesn’t load. While international service seems unlikely due to language support issues, I thought we might be able to try it out in Korea and English-speaking regions if the performance is decent. If anyone has a trial OpenAI-compatible SolOpen2 API and key, or knows a way to top up and use it, I’d appreciate it if you could let me know.

김용태 (_) 26 days ago

2
⚙️

Upstage Console

Completed

모아보기 보면 column명에 upscore라고 되어 있는 것으로 확인했습니다.

모아보기 보면 column명에 upscore라고 되어 있는 것으로 확인했습니다. 근데 이게 gold 기준으로 채점할 때, 밑의 레포 로직(upscore)으로 계산하는건가요?? 아니면 cell 단위로 채점하는건가요?? https://github.com/UpstageAI/UpScore

wooseoksong 29 days ago

1
💡

Upstage Studio

Completed

API로 Document parser를 모델을 쓰고 있습니다. job_id가 너무 많아지면 KILL할 수 있는 방법이 없나요?

현재 파이썬으로 requests 라이브러리를 사용하고 있습니다. Document Parser의 API로 호출했을 때 request_id로 KILL을 하거나 Cancel 하는 방법이 있나요? 공식 페이지를 찾아봤을 때 관련 내용이 없어서 문의드립니다. 빠른 답변 부탁드립니다.

정수민 about 1 month ago

5
💡

Upstage Studio

Completed

현재 classify 할때 문서종 목록이 추가한 순서대로 고정되어 있어, 중간에 새 문서종을 추가하면 원하는 위치에 넣을 수 없고 항상 맨 아래

현재 classify 할때 문서종 목록이 추가한 순서대로 고정되어 있어, 중간에 새 문서종을 추가하면 원하는 위치에 넣을 수 없고 항상 맨 아래에 붙는데, 문서종 항목을 드래그 앤 드롭으로 자유롭게 재정렬할 수 있으면 좋을 것 같습니다!

[조구만 루키] 길연진 about 2 months ago

2
💡

Upstage Studio

Planned

Extract 결과 중에 수정할 필요가 없는 정답인 값을 gold 결과로 일괄 반영할 수 있으면 좋을 것 같아요!

Extract 결과 중에 수정할 필요가 없는 정답인 값을 gold 결과로 일괄 반영할 수 있으면 좋을 것 같아요! 예를 들어 현재는 정답이더라도 노란 체크표시(confidence score 1.0)가 뜨면 일일히 "정답으로 표시"를 하나씩 클릭해서 반영하고 있습니다. 항목이 많아질 경우 일괄 반영 (체크 박스)할 수 있으면 어떨까 싶어 의견드립니다 :)

[조구만 루키] 길연진 about 2 months ago

1
💡

Upstage Studio