Weibo's new open supply AI mannequin VibeThinker-1.5B outperforms DeepSeek-R1 on $7,800 post-training price range

One other day in late 2025, one other spectacular consequence from a Chinese language firm in open supply synthetic intelligence.

Chinese language social networking firm Weibo's AI division lately launched its open supply VibeThinker-1.5B—a 1.5 billion parameter massive language mannequin (LLM) that may be a fine-tuned variant of rival Chinese language tech agency Alibaba's Qwen2.5-Math-1.5B.

It's out there now at no cost obtain and utilization by researchers and enterprise builders—even for industrial functions—beneath a permissive MIT License on Hugging Face, GitHub and ModelScope, with a technical report on open entry science publishing web site arxiv.org.

And but, regardless of its compact measurement, VibeThinker-1.5B achieves benchmark-topping reasoning efficiency on math and code duties, rivaling or surpassing fashions lots of of occasions its measurement, even outperforming Chinese language rival DeepSeek's famed R1 that went viral in the beginning of this 12 months—a 671-billion parameter mannequin—on formal reasoning benchmark.

It additional eclipses Mistral AI's Magistral Medium and holds its personal in opposition to Anthropic's Claude Opus 4 and OpenAI's gpt-oss-20B Medium, all whereas requiring a fraction of the infrastructure and funding.

It additionally does so having been post-trained on a price range of merely $7800 USD for compute assets (3900 GPU hours on Nvidia H800s) — far lower than the tens, and even lots of, of hundreds of {dollars} usually required to fine-tune fashions of comparable or bigger scale.

Recall this isn’t the full value of the mannequin's improvement, nevertheless: LLMs are educated in phases. First comes pre-training, when the mannequin learns fundamental language construction and normal information by predicting the subsequent phrase throughout huge quantities of textual content from the web, books, and articles. This offers it fluency however not a lot sense of learn how to observe directions or maintain a dialog

Publish-training comes subsequent, utilizing a lot smaller, higher-quality datasets—usually collections of instance questions, prompts, and expert-written solutions—to show the mannequin learn how to reply helpfully, purpose via issues, and align with human expectations. Nonetheless, Weibo's post-training value effectiveness on VibeThinker-1.5B is noteworthy and needs to be counseled.

The open-source launch upends assumptions about parameter scale, compute depth, and the minimal viable measurement for high-performance LLMs.

A Completely different Coaching Strategy: Spectrum-to-Sign

VibeThinker-1.5B owes its efficiency to not scale, however to the coaching framework behind it: the Spectrum-to-Sign Precept (SSP).

As an alternative of optimizing a mannequin purely for single-answer correctness (Cross@1), the SSP framework decouples supervised fine-tuning (SFT) and reinforcement studying (RL) into two distinct phases with totally different targets:

• SFT (“Spectrum Part”): The mannequin is educated to maximise variety throughout potential appropriate solutions, bettering its Cross@Okay rating. This builds a variety of believable answer paths.

• RL (“Sign Part”): A second-stage reinforcement studying system (referred to as MaxEnt-Guided Coverage Optimization, or MGPO) is used to determine and amplify probably the most appropriate paths from this various answer pool. MGPO prioritizes issues the place the mannequin is most unsure, utilizing entropy-based weighting to focus studying.

The authors argue this separation permits small fashions to discover reasoning area extra successfully—attaining sign amplification with out counting on huge parameter counts.

VibeThinker-1.5B makes a compelling case that the trade’s reliance on parameter scaling as the one route to raised reasoning efficiency could also be outdated.

By adopting a diversity-first coaching pipeline, WeiboAI has proven that smaller, extra accessible fashions can match and even outperform billion-dollar programs in logic-heavy duties.

The low useful resource footprint is among the many most vital elements of VibeThinker-1.5B. At beneath $8,000, the post-training value is 30–60x decrease than fashions like DeepSeek R1 and MiniMax-M1, which value between $294K and $535K to coach.

Efficiency Throughout Domains

Regardless of its small measurement, VibeThinker-1.5B delivers cross-domain reasoning that outpaces many bigger open-source and industrial fashions:

VibeThinker was benchmarked in opposition to each reasoning-centric fashions (Magistral, Claude, OpenAI o3-mini) and non-reasoning LLMs (GPT-4.1, Kimi K2, DeepSeek V3). Throughout structured reasoning benchmarks, the mannequin persistently outperformed non-reasoning fashions, no matter measurement:

• On AIME24 (math), it beat Kimi K2 (1.09T) by over 10 factors (80.3 vs. 69.6).

• On LiveCodeBench v6, it surpassed Claude Opus 4 (51.1 vs. 47.4).

• On GPQA, it scored beneath GPT-4.1 and Claude, however nonetheless doubled its base mannequin (from 16.4 to 46.7).

This helps the authors’ declare that measurement will not be the one path to reasoning functionality—with correct coaching design, smaller fashions can attain and even exceed the efficiency of far bigger programs in focused duties.

Notably, it achieves parity with fashions lots of of occasions bigger on math and code, although it lags behind on the whole information reasoning (GPQA), the place bigger fashions keep an edge.

This implies a possible specialization trade-off: whereas VibeThinker excels at structured logical duties, it has much less capability for wide-ranging encyclopedic recall, a identified limitation of smaller architectures.

Steerage for Enterprise Adoption

The discharge consists of really helpful inference settings (temperature = 0.6, top_p = 0.95, max tokens = 40960).

The mannequin is sufficiently small to be deployed on edge gadgets, together with cell phones and vehicle-embedded programs, whereas inference prices are estimated to be 20–70x cheaper than with massive fashions.

This positions VibeThinker-1.5B not simply as a analysis achievement, however as a possible basis for cost-efficient, regionally deployable reasoning programs.

Weibo’s Technique and Market Place

Weibo, launched by Sina Company in 2009, stays a cornerstone of China’s social media ecosystem. Typically described as China’s model of X (previously Twitter), the platform blends microblogging, multimedia content material, and trending-topic options with a regulatory surroundings formed by tight authorities oversight.

Regardless of counting 600 million month-to-month lively customers (greater than twice that of X), traders aren’t optimistic about its promoting income progress potential within the close to time period, and Weibo is navigating intensifying competitors from video-first platforms like Douyin, that are drawing youthful customers and growing time-spent elsewhere.

In response, Weibo has leaned into creator-economy monetization, live-streaming, and vertical video—including instruments for influencer engagement, e-commerce integration, and richer analytics for manufacturers.

The platform’s position as a digital public sq. additionally makes it a spotlight of regulatory scrutiny. Chinese language authorities proceed to use strain on points starting from content material governance to information safety. In September 2025, Weibo was among the many platforms cited in official warnings, highlighting its ongoing publicity to coverage dangers.

Weibo’s push into AI R&D—exemplified by the discharge of VibeThinker-1.5B—alerts a shift in ambition. Past being a media platform, Weibo is positioning itself as a participant within the subsequent section of Chinese language AI improvement, utilizing its capital reserves, person conduct information, and in-house analysis capability to pursue adjoining technical domains.

What It Means for Enterprise Technical Determination Makers

For engineering leaders and enterprise AI groups, VibeThinker’s launch has sensible implications for every little thing from orchestration pipelines to value modeling.

A 1.5B-parameter mannequin that outperforms 100x bigger fashions on math and programming duties doesn’t simply save compute—it shifts the architectural steadiness. It allows LLM inference on constrained infrastructure, reduces latency on the edge, and lowers the barrier to entry for purposes that in any other case would have required API entry to closed, frontier-scale fashions.

That issues for enterprise ML leads attempting to deploy reasoning-capable brokers inside current programs, or for platform homeowners tasked with integrating LLMs into automated workflows.

It additionally speaks to these working reinforcement studying from human suggestions (RLHF) pipelines or managing inference optimization throughout hybrid cloud environments.

The mannequin’s post-training methodology—notably its entropy-targeted reinforcement studying strategy—presents a roadmap for groups trying to refine smaller checkpoints as a substitute of counting on large-scale pretraining.

VibeThinker’s benchmark transparency and information decontamination steps additionally tackle one other rising precedence in enterprise AI: auditability. Whereas its efficiency on general-knowledge exams nonetheless trails massive frontier fashions, its task-specific reliability makes it a pretty candidate for managed environments the place correctness issues greater than protection.

Briefly, VibeThinker-1.5B isn’t only a analysis milestone—it’s a powerful candidate for sensible enterprise use, deployment and learnings. It suggests {that a} new class of compact, reasoning-optimized fashions is viable for enterprise use instances that had been beforehand the area of far bigger programs. For organizations attempting to steadiness value, latency, interpretability, and management, it’s a very good new choice to the lengthy, rising checklist of Chinese language open supply choices.