Upstage 发布 Solar Mini 4 推理模型，35B 总参数 3B 激活

All articles

September 30, 2026

Upstage has released Solar Mini 4, a new proprietary reasoning model. Upstage reports 35B total and 3B active parameters, setting a new Pareto optimal point on Intelligence Index vs. Active Parameters for models under 3B active parameters. It also scores 16 points higher than Upstage's previous-generation flagship, Solar Pro 3 (8), and cuts per-token pricing by a third to $0.10/$0.40 per 1M input/output tokens.

Key results:

➤ Strong for its reported active parameter size: Solar Mini 4 scores 6 points higher than Qwen3.6 35B A3B (Reasoning), which has the same 3B active parameters, and 1 point higher than Nemotron 3 Ultra, which has 55B active. As a proprietary model, its size cannot be independently verified.

➤ Long context reasoning is a relative strength: It scores 83% on AA-LCR v1.1, matching MiniMax-M3 and GPT-6 Luna (max), and ahead of Gemini 3.8 Flash (high) and GPT-6 Astra (max) at 81%. It also scores 48% on SciCode, ahead of MiniMax-M3 and Inkling (xhigh) at 47%.

➤ Fast output, but slow tasks: Solar Mini 4 generates 208 tokens/s as of launch date, faster than GPT-6 Luna (max) at 152 tokens/s. However, it uses 88k output tokens per Intelligence Index task, so it averages 7.1 minutes per task.

➤ Agentic coding is a relative weakness: It scores 1% on Terminal-Bench 4.0 and 22% on AutomationBench-AA. On agentic knowledge work, it scores 1072 Elo on GDPval-AA and 872 Elo on AA-Briefcase, close to Inkling (xhigh).

➤ Low knowledge accuracy, but also relatively high non-hallucination rate: Solar Mini 4 scores -11 on AA-Omniscience with 18% accuracy. It abstains on about half of the questions, and scores 64% on non-hallucination rate, higher than Inkling (xhigh) at 32% and GPT-6 Luna (max) at 23%.

➤ License: Proprietary, with weights not released

➤ Parameters: 35B total, 3B active (reported by Upstage)

➤ Modalities: Text input and output only

➤ Knowledge cutoff: February 2026

➤ Pricing: $0.10/$0.40/$0.01 per 1M input/output/cache hit tokens

Upstage reports that Solar Mini 4 has 35B total and 3B active parameters, and at the claimed size it sets a new Pareto optimal point for Intelligence Index vs. Active Parameters. Solar Mini 4 is a proprietary model, so its weights are not public. Qwen3.6 35B A3B (Reasoning) scores 18 with the same 3B active parameters, while K2 Horizon MoVA 36B A4B scores 25 with 4B active. Solar Pro 3, Upstage's previous-generation flagship, scores 8 with 12B active parameters.

Uncached input makes up $0.30 of Solar Mini 4's $0.36 cost per Intelligence Index task. Agentic tasks such as Terminal-Bench 4.0 and AA-Briefcase resend the growing conversation each turn, so cache hits matter. In our measurements, 48% of Solar Mini 4's repeated context was served from cache vs. 99% for GPT-6 Luna (max), which costs $0.07 per task. Reasoning and answer tokens add only ~$0.04 per task.

Solar Mini 4 uses 88k output tokens per Intelligence Index task, more than Claude Fable 5.1 (max with fallback) at 78k. 72k of these are reasoning tokens. It uses ~2.5x as many output tokens as Inkling (xhigh) and ~5x as many as Gemini 3.5 Flash-Lite, which score 25 and 22 respectively. At $0.40 per 1M output tokens, this adds little to cost, but it makes tasks longer.

Solar Mini 4 generates 208 tokens/s, but heavy token use makes it slower per task than GPT-6 Luna (max). It averages 7.1 minutes of decode time per Intelligence Index task, vs. 5.8 minutes for GPT-6 Luna (max) at 152 tokens/s and 2.8 minutes for Inkling (xhigh) at 183 tokens/s.

Solar Mini 4's AA-Omniscience score of -11 reflects low knowledge accuracy rather than a low non-hallucination rate. It answers 18% of questions correctly, among the lowest of the models compared, and abstains on about half. Its non-hallucination rate of 64% is well ahead of Inkling (xhigh) at 32% and GPT-6 Luna (max) at 23%.

Artificial Analysis Decision Index v1.0

The Artificial Analysis Decision Index benchmarks decision models, a new category of models that answer structured decisions in a single call with a probability for each option. Version 1.0 combines four evaluations across classification, tool calling, knowledge, hallucination, and reasoning.

October 13, 2026

Announcing AA-3D: a benchmark for AI models building interactive 3D experiences

AA-3D is a new benchmark of how well AI models build interactive 3D experiences. Each model turns a written brief into a working three.js scene in an offline sandbox, and our recruited panel compares the scenes in blind pairwise votes.

October 12, 2026

Announcing FrontierCode on Artificial Analysis: can AI models write code maintainers would merge?

Cognition's FrontierCode 1.1 tests whether AI models make code changes that maintainers of real open source repositories would merge. Leading models still fail more than 40% of the time.

October 12, 2026

Artificial Analysis

Get notified about new articles

About; Methodology; Contact; Articles; Careers