Upstage

Upstage ships Solar Pro 4, aimed at multi-step work rather than single answers: 512K context with 128K output, a reasoning-effort dial, and scores of 57 on Terminal-Bench v2.1 and 71 on AA-LCR. Trained on graded office-task deliverables in Korean, English and Japanese.

Upstage launches Solar Mini 4, a 35B MoE activating 3B parameters per token with a 512K context window, vendor-reported agent benchmark results, and API pricing at $0.10 input and $0.40 output per million tokens.

A sequential fine-tuning recipe for math reasoning: chain-of-thought training to establish decomposition, followed by program-of-thought training that hands computation to a code interpreter, plus a cognitive retention step. Reported as state-of-the-art on GSM8K and MATH over larger models.

Solar Open 2 arrives as open weights under a commercial license: a 250B-parameter MoE activating 15B per token, a 1M-token context, and quantized deployment on two H200s. Trained on search, tool-calling, coding and document work, with a full technical report on architecture and evaluation.

Upstage extends Document Parse beyond text and tables to charts, emitting HTML instead of asking an LLM to read the picture. Evaluated in two stages, on isolated chart images and on full document pages with question answering, against parsers from Google and Microsoft.

Solar Pro and Solar Mini get a Korean-focused update, with Pro gaining 7.6 points on Upstage's internal fluency metric and jumping 71.8 to 81.0 on Ko-MT-Bench. Both versions are live on the Upstage console, AWS Marketplace and on-prem, with comparisons to Exaone and HCX included.

Upstage promoted Information Extract to a production REST API that returns typed JSON matching a caller-supplied schema. It handles tables, checkboxes, rotated content, and multi-page layouts without templates or fine-tuning, and prices per page rather than per token.

Upstage open-sourced Evalverse, an evaluation library that pulls lm-evaluation-harness and FastChat in as git submodules so upstream changes flow through, and adds a Slack bot for triggering benchmark runs and generating comparison reports without writing code.

Upstage and Karakuri released Syn Pro, a sub-32B Japanese LLM ranked first among locally trained Japanese models and in the global top 20 on the Nejumi leaderboard. It deploys on-prem, in private cloud, or on your own GPUs, with a reasoning-budget control for speed versus depth.

Most table recognition models emit HTML tags plus bounding boxes, then fight to match them against OCR output. TFLOP drops the boxes for a layout pointer that maps cells onto text regions directly, adds span-aware contrastive supervision for merged cells, and leads on the public TSR benchmarks.

Upstage launched Solar Mini 4, a 35B mixture-of-experts model with 3B active parameters built for high-volume agent work. It is on Upstage Console, Solar Chat and OpenRouter, with on-prem support and 70% off pricing until October 10.

Extremely tall product images, standard in Korean, Japanese and Chinese e-commerce, break most document parsers. Upstage's document-parse-250404 handles them with a claimed 38.6% accuracy improvement and returns HTML intended as the input stage for LLM key-value extraction.

A parser built for RAG pipelines: PDFs and scans converted to HTML or Markdown with layout and table structure preserved. Upstage claims 100 pages per minute and publishes DP-Bench, its TEDS-based comparison against Unstructured, LlamaParse, Textract and the Google and Microsoft parsers.

Upstage's Document Parse added stitching for tables spanning multiple pages, handling of 90, 180, and tilted scans, roughly 2x faster processing of long vertical images, HWP and HWPX conversion, and better table recognition in dense form layouts.

Upstage is offering its Solar Pro 4 model free of charge inside the open-source Cline coding agent through September 15, positioning it for long, multi-turn tool-calling tasks across a full repository.

Upstage releases Solar Pro Preview under an open license for commercial use, positioning it against other single-GPU models on MMLU-Pro and IFEval and claiming parity with multi-GPU models like Llama 3.1 70B. VRAM comparisons are stated with the measurement method, at 4k context.

Upstage's March Solar Pro 3 update discloses the architecture behind the jump: a 102B mixture-of-experts with 12B active parameters, trained with an in-house RL framework, roughly doubling tool-calling and terminal-workflow benchmark scores while holding the API interface and throughput steady.

Upstage open-sources Dataverse, a Python ETL library for large-scale LLM training data. It wraps Spark so collection, filtering, deduplication and cleaning run distributed without writing Spark directly, and the post lays out the architecture and the fairness motivation behind it.