As small AI models that can be installed on PCs rapidly evolve while delivering solid performance, attention is growing on whether they can serve as alternatives to large AI models that run on high-capacity computers.

Major technology companies have recently been rolling out small models that can be installed on PCs and deliver large-model-level performance, targeting companies focused on cost savings and privacy.

Alibaba has released as open source a multimodal AI model, Qwen3.8-27B, optimized for local running. Qwen3.8-27B is sized to run locally on high-spec MacBook Pro or Mac Studio machines. But it is competitive with models several times larger in coding and office-work benchmarks. Besides Alibaba, China’s DeepSeek and, in the United States, Meta have recently unveiled PC-installable models with relatively strong performance. Nvidia is also speeding up efforts to expand its lineup of AI models aimed at local environments.

ㆍAlibaba launches AI model that can be installed on MacBooks...Opus 4.6 Max-level performance draws attention ㆍPC-installed LLMs evolve rapidly...becoming an AI battleground

The government will soon release the results of a second evaluation of four elite teams participating in its "independent AI foundation model" project. It will combine benchmarks, expert assessments and user evaluations involving 200 members of a public evaluation group to select 3 of the 4 teams to advance to the next stage. The public evaluation ran through the 11th. Ahead of the announcement, rival companies are also stepping up promotions highlighting their models’ performance.

ㆍDokpamo’s 4 companies complete preparations for second evaluation...different 'scale-up' strategies draw attention ㆍHow will Dokpamo’s second round evaluate agentic AI?..."Tool calls allowed in expert evaluation" ㆍ[Exclusive] How will the Dokpamo public evaluation work?..."Mix of common questions and free-form trials" ㆍSKT’s independent AI model A.X K2 posts world-class results in AI math-reasoning evaluation ㆍMotif’s 'Motif 3' scores 47 points on AAII...No. 9 globally and No. 1 in Korea

In the AI market, the weight of costs also appears to be growing. Of tokens moving through the AI model routing platform OpenRouter, 84 percent are handled by models that are not top-performance models. The 6 models that handle most tokens deliver performance at 77 percent of top-performance models, while costing only 2.5 percent of the price of Claude Fable 5. Major U.S. AI companies are also speeding up price cuts as part of efforts to keep Chinese AI companies in check.

ㆍ[Tech Insight] The changing economics of AI models..."It will become increasingly difficult to recoup costs for cutting-edge models" ㆍGlobal AI market gains momentum in shifting to 'price-centric' restructuring...major U.S. companies also mount a price-cut offensive

The article also summarizes recent moves by companies in Korea and abroad around AI.

Hancom is accelerating AI conversion for its existing office customers while expanding its global business based on the commercialization of an agentic OS. Hancom will pursue its AI business in the second half through three pillars: AI conversion of existing customers, commercialization of an agentic OS, and expansion into the global physical AI market. Samsung SDS said a total of 9 government ministries and agencies, including the Ministry of Economy and Finance and the former Ministry of Planning and Budget, as well as the Ministry of Patriots and Veterans Affairs, the Ministry of Trade, Industry and Energy, the Ministry of Employment and Labor, the Financial Services Commission and the Overseas Koreans Agency, decided to adopt its AI collaboration solution, Brity Works. Cloud and AI management company Cloucus was selected as an OpenAI Select Partner under the OpenAI Partner Network. MegazoneCloud was selected by Amazon Web Services' Generative AI Innovation Center (GenAI IC) as a systems integration partner for the intelligent AI assistant "Amazon Quick."

ㆍHancom accelerates AI-focused strategy in second half...agentic OS beta to be unveiled in the third quarter ㆍSamsung SDS AI collaboration tool 'Brity Works' to be adopted by 9 central government ministries ㆍCloucus cooperates with OpenAI...accelerates push into enterprise market ㆍMegazoneCloud selected as SI partner for 'Amazon Quick'...targets enterprise AI agent market

Data foundry company Boundfor added an AI token spending management service, Pareto, to its AI data operations platform, DroPai. B2B financial AI company Webcash will launch a B2B vibe-coding integrated platform for working-level practitioners, AXPort, in September. Naver Cloud unveiled 3 new Naver Works features: "Naver AI Search" within the administrative network, the document collaboration and knowledge management service Page, and the AI agent-building solution EASY (Enterprise Agent Solution for You).

ㆍBoundfor launches AI token spending management service 'Pareto' ㆍWebcash to launch B2B vibe-coding platform 'AXPort' in September ㆍNaver Cloud expands 'Naver Works' AI features...adds search and agents

OpenAI unveiled a new "Ultrafast" mode that boosts GPT 5.6 Sol processing speed by up to 14 times. Ultrafast mode can generate up to 750 output tokens per second. OpenAI released a preview version of its ChatGPT desktop app for Linux to global users.

ㆍOpenAI unveils GPT 5.6 Sol acceleration mode 'Ultrafast' ㆍOpenAI releases ChatGPT desktop app for Linux

Google launched a new AI model, Gemini 3.7 Flash, specialized for coding and automation tasks. At its "Made by Google 2026" event, Google unveiled the Pixel 11 series, Pixel Watch 5 and the "Pixel Tag" item tracker, and expanded Gemini-based features across devices. Monthly active users of the Google Gemini app have surpassed 1,000,000,000.

ㆍGoogle unveils 'Gemini 3.7 Flash'...timing for Gemini 3.5 Pro release remains undecided ㆍGoogle unveils new Pixel 11 smartphone series...also launches 'Pixel Tag' to find lost items ㆍGoogle unveils sign-language-understanding AI model 'SL2T'...first applied to Pixel 11 ㆍGoogle Gemini app tops 1,000,000,000 monthly users

Microsoft will integrate its consumer Copilot app and Microsoft 365 Copilot app and remove some AI features that failed to deliver results. According to TechCrunch on the 13th local time, the overhaul aims to simplify the Copilot experience in line with a trend in which personal and work AI use overlap.

ㆍMicrosoft integrates scattered Copilot apps...removes AI features that failed to deliver results

Chinese AI startup Z.ai unveiled GLM-5.3, an AI model specialized for cybersecurity. The company said GLM-5.3 greatly improved performance over the previous model in vulnerability detection, exploit analysis and multi-step security tasks. It plans to 공개 its weights soon. Meta, which shifted its center of gravity to a closed approach after competing with open-source AI models, is again strengthening its open-source AI model strategy. Meta will 공개 the weights for its top-tier AI model, Muse Spark 1.2, and will also release a new open-source model for consumer devices.

ㆍChina’s open-weight models also target security...Z.ai challenges Mithos with GLM 5.3 ㆍMeta returns to the open-weight AI market...open-sources even its top version

Data and AI company Databricks, according to the company, said its annualized revenue has surpassed $7 billion. Databricks grew more than 80 percent in the second quarter from a year earlier. Databricks also completed a $5 billion strategic investment round valuing the company at $190 billion. IBM is expanding its enterprise AI business with OpenAI. The two companies announced a partnership to provide OpenAI models and tools to enterprise customers.

ㆍDatabricks annualized revenue tops $7 billion ㆍIBM partners with OpenAI...expands enterprise AI business

Nvidia’s moves to develop its own open-source AI models are accelerating. It is developing and distributing its own open-source AI models as part of efforts to expand hardware demand, but some see this could place Nvidia in a delicate position where it may have to compete with AI companies that buy its hardware.

ㆍ[Tech Inside] Why is Nvidia trying to develop even large AI models itself? ㆍNvidia’s development of open-source AI models gains momentum..."Developing Nemotron4 with 1 trillion parameters"

A so-called "SaaSpocalypse" is resurfacing, a view that AI will eliminate or significantly shrink the space for software-as-a-service (SaaS) companies. The debate appears to be intensifying, especially after second-quarter earnings releases by major global SaaS companies.

ㆍRound 2 of the 'end of SaaS' debate...AI-maladapted SaaS faces a crisis narrative