Something large has happened - a acheronian equine has emerged among China's apical home ample exemplary developers.
A caller company, pinch its first exemplary having only 27B parameters, took 2nd spot wide successful the CAICT MCP specialized test.
Ranked up of it is the slayer limb Liang Wenfeng has kept nether wraps for astir a year: DeepSeek-V4-Pro, boasting a parameter count of 1.6 trillion.
The quality betwixt the 2 is simply a specified 1.3 percent points.
This acheronian equine is the StartLux-V1.0-27B-Preview, from StartLux (formerly Yuandian Xinghui).
Looks a spot unfamiliar, doesn't it? Don't worry, its laminitis and CEO is an aged acquaintance: Chen Danyan.
Known arsenic the "godfather" of programmers successful the net era, his achievements are a matter of nationalist record:
Shanda Network's co-founder and the caput of Lian Shang Network, 1 of China's earliest programmers to present the conception of "shareware"
After a decade of retirement, he has made a comeback, this clip betting connected section models.
This has to do pinch Chen Dawei's caller uncommon nationalist appearance.
At the 18th day reunion of Shanda Innovation Institute, he publically declared "eight non-consensus views for the AI era," 4 of which halfway connected section models.
Local models will wholly destruct the unreality market, catching up pinch Claude successful 3 years and occupying 80% of the market. The exemplary title based connected parameters is going to beryllium outdated...
StartLux is the champion practice of his idea.
The company's business has not followed the manufacture trend, alternatively focusing connected the commercialization of mini and beautiful section models, making it the world's first genuinely section exemplary institution successful the existent sense.
As the first market-oriented scorecard, StartLux-V1.0-27B-Preview does not trust connected the unreality and tin tally straight connected consumer-grade PCs.
In different words, this section model, which is astir 60 times smaller, has Agent capabilities that tin lucifer those of a trillion-level cloud-based flagship model.
What justification is location for this?
Small Model Achieves Big Results, 27B Outperforms 1.6T
Before the reply is revealed, let's return a look astatine who the comparison is being made to.
It's often said that cipher remembers the 2nd place, unless the first is DeepSeek.
Moreover, the spread is minimal, making it good worthy discussing.
The results travel from the charismatic institution, China Academy of Information and Communications Technology's trustworthy AI ample exemplary benchmark trial MCP typical item, which sets six types of tasks astir existent exertion scenarios:
Location navigation, web search, browser automation, financial analysis, codification repository management, 3D design.
An further broad appraisal will beryllium added, focusing connected evaluating Agent's multi-tool collaboration, analyzable task execution, and relationship successful real-world environments.
In elemental terms, MCP-Universe doesn't measure models based connected their responses, but alternatively connected whether they tin really get things done. This is besides the astir basal facet of judging an Agent's quality.
The results showed that StartLux-V1.0-27B-Preview had a broad people of 39.25%, ranking second.
DeepSeek-V4-Flash-0731, pinch complete 284 cardinal parameters, and Step-3.7-Flash, astatine 198 billion, way DeepSeek-V4-Pro by conscionable 1.3 percent points.
With the aforesaid 27B parameter scale, StartLux besides surpasses Qwen 3.6 by 5.34 percent points.
It besides excelled successful individual subjects, ranking first successful location navigation, financial analysis, and browser automation, pinch its different sub-items besides ranking high.
Let's return a look astatine 2 lawsuit studies, putting information aside.
The first mobility is astir a two-year Microsoft banal investment, pinch Claude Sonnet 4.6 arsenic the competing topic.
Claude's reply is: $47,254, 89.02%.
Video link: https://mp.weixin.qq.com/s/365CtdgGYFlEKNDICtoCfg
StartLux gave: $47,499.09, 90.00%.
It whitethorn look similar, but successful the financial industry, a mini quality tin lead to tremendous losses.
Careful introspection of the 2 models' reasoning processes shows that, owed to missing earthy data, Claude Sonnet 4.6 misidentified January 8, 2025 arsenic a non-trading time and alternatively calculated the erstwhile day's closing price.
Under the aforesaid circumstances, StartLux retrospectively reviews the original information and verifies the marketplace trends astir the target day to corroborate the meticulous closing value earlier completing the calculation and generating visualization.
Ultimately, the conclusion reached by StartLux proved correct, and it was afloat verifiable and traceable.
The 2nd task was much straightforward: some models were asked to hunt for formation tickets successful a browser astatine the aforesaid time.
I'm incapable to unfastened a browser aliases interact pinch unrecorded websites for illustration Google Flights. I tin only process matter and don't person real-time browsing capabilities. To find this formation yourself, here's what you'd do: 1. Go to google.com/flights 2. Enter Singapore (SIN) → Beijing 3. Select one-way, group departure day to 5 days from today 4. Filter by "Nonstop" and "Economy" 5. Check the results — statement that flights to Beijing whitethorn onshore astatine either Capital (PEK) aliases Daxing (PKX). Exclude immoderate Daxing arrivals. 6. Compare prices and prime the cheapest nonstop action landing astatine PEK. If you'd like, I tin thief you deliberation done emblematic value ranges, hose options connected this route, aliases wide tips for uncovering inexpensive flights.
Among them, StartLux-V1.0-27B-Preview completed the hunt successful astir 95 seconds, uncovering an Air China summons priced astatine $299.
By contrast, Claude Sonnet 4.6 required much screenshot confirmations erstwhile navigating day selection, popup dismissal, and select menus, and moreover accidentally triggered the clip select sheet astatine 1 point.
More than 200 seconds later, it gave a lowest quote of $556. The value was higher, and the hunt clip was still doubly that of StartLux.
In particular, successful position of operational pathways, StartLux is overmuch much concise, requiring only 12 steps, whereas Sonnet requires a afloat 21 steps.
This is capable to exemplify that, successful Agent tasks, the standard of parameters is nary longer the only decisive variable.
New variables are being introduced done post-training.
Cutting Prices, Not Capabilities
StartLux-V1.0-27B-Preview was not trained from scratch.
It is besides based connected Qwen3.6-27B, but the last trial people is importantly higher than Qwen, and the logic lies successful the task information and automated post-training methods.
In elemental terms, StartLux trains a much tin Agent, pinch training information focusing connected reinforcing abilities specified arsenic task understanding, instrumentality selection, parameter construction, multi-step execution, position checking, and consequence verification.
The exemplary needs to study not only to output the last text, but besides erstwhile to invoke which tool, really to set erstwhile a instrumentality returns an exception, and nether what circumstances it tin state the task complete.
Further training will push this process moreover further.
The squad has independently developed a brand-new, multi-dimensional, verifiable, and scalable exemplary loop optimization technology, which uses the AI-trained AI (Auto Research) method to alteration the exemplary to autonomously execute tasks successful a real-world instrumentality situation and continuously set its strategy based connected biology feedback.
For instance, the financial study lawsuit mentioned earlier, which involves backtesting and revision, arsenic good arsenic the constraint recognition and way action successful browser tasks, are the astir intuitive manifestations of post-training.
According to charismatic information, StartLux-V1.0-27B-Preview is besides the country's first section Agent exemplary to complete post-training utilizing the Auto Research method.
This does not mean that the Scaling Law is invalid.
Large parameter unreality models are still the mainstream prime astatine present, but StartLux has besides fixed a clear signal: this is not the only solution.
In the words of Chen Danyan:
Scaling Law is simply a "passing fairy," without it, AI cannot return off, but it has simply passed done the improvement way of AI and its early is not needfully tied to it.
The manufacture has besides go alert of this issue, and the method way for ample models is presently showing a inclination of chopped divergence:
On 1 hand, location are the die-hard believers successful "more powerfulness leads to miracles," pinch parameters scaling from tens of billions to hundreds of billions, and past to trillions, while training costs besides surge exponentially;
On the different hand, the Agentic appraisal system, represented by Harness, has quickly gained popularity, pinch an expanding number of experts and scholars explicitly advocating for "less is more".
To paraphrase Wang Yangming, the unity of knowledge and action intends higher-quality action is what genuinely matters. StartLux offers different illustration of simplifying models.
This tin besides explicate why StartLux insists connected section models.
Once the exemplary is connected a PC, for semipermanent tasks, its costs building tin displacement from continuously accumulating cloud-based token fees to much controllable instrumentality computing powerfulness and energy consumption, while the exemplary tin besides amended understand the user's agelong context, achieving much personalized goals.
It's not conscionable StartLux, arsenic Meta, Google, and NVIDIA person besides precocious been expanding their finance successful section mini exemplary development.
However, astir of these are still successful the experimental shape aliases cater to niche groups, pinch only 1 institution focusing connected section exemplary commercialization.
StartLux besides stated that they will steadily beforehand their ain instauration exemplary training and research caller architectures specified arsenic diffusion-based connection models, and arsenic agelong arsenic they are connected the correct path, the early is promising.
StartLux has taken complete the section model, and since this is simply a non-consensus route, those astatine the helm request to beryllium 2 types of people: those who situation to return bets and those who tin present results.
StartLux's all-star squad exemplifies this, pairing entrepreneurs pinch scientists successful a formidable alliance.
StartLux laminitis Chen Danyan
CEO Chen Danyan had antecedently been concisely introduced arsenic a serial entrepreneur and 1 of China's first-generation programmers, who roseate to fame astir the aforesaid clip arsenic Zhang Xiaolong and Lei Jun, and started his business ventures successful the aforesaid era arsenic Ma Yun and Ma Huateng.
He was 1 of the first group successful China to present the conception of "shared software" and co-founded Shanda Network pinch his relative Chen Tianqiao, arsenic good arsenic the Shanda Innovation Institute, a cradle of net innovation, and was besides instrumental successful incubating the nationally celebrated merchandise "WiFi Master Key".
It tin beryllium said that he is highly acquainted pinch Chinese marketplace users and products, and section models are besides his comfortableness zone.
StartLux co-founder Guo Quanwei
The personification responsible for implementing the method roadmap is StartLux co-founder and CTO, Guo Quanwei.
Kuo Chuan-wei holds a Ph.D. successful Computer Science and Engineering from National Yang Ming Chiao Tung University, pinch investigation areas covering locally deployed ample models, Agentic AI, AI for Science, AI for Finance, and privacy-preserving instrumentality learning.
Before joining StartLux, he served arsenic the Chief Algorithm Scientist astatine AI Science institution Huanliang Technology, and earlier worked astatine the Industrial Technology Research Institute of Taiwan, wherever he developed information privacy, information de-identification, and privacy-preserving instrumentality learning.
He is besides a recipient of the 2024 TAAI Best Paper Award and holds data-privacy-related invention patents arsenic the first named inventor.
Another co-founder of StartLux, Luo Yongxiang, is the erstwhile Managing Director of Morgan Stanley Asia.
Chen Danyan understands products and users, Guo Quanwei has agelong studied section models, Agents, and information privacy, and Luo Yongxiang is successful complaint of trading and finance financing. This operation is highly suited to StartLux and will besides thief thrust StartLux's semipermanent development.
As for what the squad yet wants to deliver, it's not conscionable a group of exemplary weights.
In StartLux's vision, section intelligent solutions should beryllium deployable pinch 1 click, akin to installing Office. Users do not request to understand quantization, GPU representation configuration, and conclusion frameworks, nor do they request to optimize them many times themselves.
The squad presently plans to motorboat its first-generation section smart solution for endeavor and individual users wrong the year.
In this light, the emergence of StartLux is by nary intends conscionable "another player" entering the scene.
It besides represents that, pursuing the emergence of cloud-based exemplary companies for illustration DeepSeek and Kimi, home section models are besides starting to capable successful the gaps.
From a rising prima successful the intelligent era to circling backmost to the first-generation programmers of the net era, the basal paradigm of models is shifting gears—but done it all, Chinese companies person kept passing the baton and pushing forward.
Official website link: https://startlux.com/
English (US) · 
Indonesian (ID) · 
