One other BIG AI from China! LongCat-Flash Chat 560B

China is again with one other LLM. This time, it isn’t Qwen, DeepSeek or Kimi within the highlight. The thrill is round Meituan’s LongCat-Flash. Identified primarily as a number one meals supply big, Meituan has shocked the open-source group with the discharge of its extremely succesful mannequin. Early benchmarks present LongCat-Flash acting on par with, and in some instances outperforming, friends like DeepSeek, Qwen and Kimi, because of its extra environment friendly structure.

On this weblog, we’ll discover LongCat-Flash, its options, efficiency, accessibility and real-world functions.

What’s LongCat-Flash?

LongCat-Flash is the most recent massive language mannequin from Chinese language tech big Meituan. It’s a 560-billion-parameter mannequin constructed on the MoE structure. A key innovation is its dynamic computation system, which prompts between 18.6B and 31.3B parameters (averaging round 27B) relying on context, effectivity, and efficiency wants. The mannequin additionally makes use of shortcut-connected MoE, permitting it to compute and talk concurrently, making it quicker and extra succesful than different fashions of comparable dimension.

The launched model, LongCat-Flash-Chat, is a basis mannequin with agentic capabilities. Whereas not multimodal, it’s a sophisticated text-based mannequin optimized for reasoning and agent-driven duties.

Key Coaching and Structure Highlights of LongCat-Flash

Meituan’s newest LLM comes with very revolutionary approaches in the best way it’s skilled and developed.

A few of its key options are:

Computational Effectivity

The mannequin is constructed to effectively take care of enormous quantities of information whereas holding the prices underneath examine. It does this by:

• Counting on “zero computation consultants,” which is a brilliant system focusing solely on the necessary components and makes use of parameters primarily based on want. 
• Utilizing a design referred to as “Shortcut linked MoE (ScMoE), which decreased the time that’s wasted between inter-machine communication. Thus, it saves time by dashing up each coaching and inference. 

Scaling Technique

The LongCat-Flash mannequin is scaled whereas prioritizing stability and reliability. That is accomplished by:

• Borrowing settings from smaller fashions, i.e., the mannequin predicts the suitable hyperparameters for the massive fashions by studying from the smaller ones. 
• Beginning the mannequin with a correctly ready “half-scale” model, which boosts its efficiency proper from the beginning. 
• Balancing gradients, controlling enormous activations, and tuning optimizers in order that the mannequin doesn’t crash or lose high quality. 
• Ensuring coaching runs the identical method every time makes it simple to breed the experiments and catch the hidden errors. 

Multi-Stage Coaching Pipeline

To make this LLM work like an agent that may purpose, work together, and remedy advanced issues, a multi-stage coaching pipeline was constructed, consisting of:

• In Pre-training, it utilized a mixture of reasoning-heavy information to make sure the bottom mannequin with drawback problem-solving capabilities. 
• In Mid coaching course of, its reasoning and coding expertise have been enhanced and its context size was prolonged to 128k tokens to permit it to recollect longer conversations.
• In Publish coaching, it leverage a multi-stage set as much as sharpen its agentic expertise and simply to get the fashions style of actually troublesome issues. Meituan’s staff developed a multi-agent synthesis framework to generate these powerful coaching issues.

The best way to Entry LongCat – Flash?

Chat

Don’t fear in case you are not in a position to create your account, you possibly can nonetheless work together with the mannequin with out creating an account. However with out an account your chats gained’t be saved, therefore, you gained’t be capable to revisit your dialog. 

Hugging Face

• Simply head right here.
• On the suitable facet, you will see the choice to “Use this mannequin”.
• Right here you will see the code to make use of this mannequin in Google Colab, Kaggle or vLLM.

For those who don’t have colab Professional then you definitely may discover it troublesome to make use of this mannequin( its a heavy LLM with over 500 B parameters.) LongCat-Flash’s API is just not but accessible. 

Working with LongCat-Flash

Now that we all know tips on how to entry this mannequin, we’ll take a look at its capabilities for 3 completely different duties: 

• Coding
• Reasoning
• Agentic Capabilities

Let’s take a look at this mannequin. 

Activity 1: Creating an HTML Web page

Immediate:

Write the HTML code for making a button that, when clicked on, generates confetti. 

Output: 

    

    

    

    





    Click on for Confetti!

    

The mannequin generated a protracted piece of code for a easy activity, but it surely labored completely. It produced a button that triggered confetti when clicked. The mannequin responds rapidly and even double-checks the code it generates.

Activity 2: Fixing a Query

Immediate

A projectile is thrown from some extent O on the bottom at an angle 45° from the vertical and with a pace of 5 √2 m/s . The projectile on the highest level of its trajectory splits into two equal components. One half falls vertically all the way down to the bottom, 0.5 s after the splitting. The opposite half, t seconds after the splitting, falls to the bottom at a distance x meters from the purpose O. The acceleration because of gravity g = 10 m/s2 . The worth of t is ________.

Output:

The mannequin used a protracted and tedious technique to calculate the worth of t. Though the query was easy, it took a number of pointless steps. Ultimately, it calculated the worth appropriately however second-guessed itself and offered an incorrect closing reply.

Activity 3: Agentic Capabilities

Immediate

Discover the most recent comedy reveals taking place in Gurgaon within the week of October 1 – 7 and discover me the reveals for which the ticket worth is the most cost effective. 

Output:

The mannequin was not in a position to fetch precise information to search out the related reveals taking place round me within the given date. It offered some hyperlinks within the reply however none of them have been working or led to an precise net web page. All its outcome was primarily based on hypothetical potentialities.

LongCat-Flash: Efficiency Benchmarks

Now that we now have seen the outcomes that the LLM generates on completely different duties, let’s have a look at the efficiency benchmarks of LongCat-Flash in comparison with the opposite prime fashions. 

• LongCat-Flash scores persistently excessive in duties like CEval and MMLU showcasing sturdy basic information expertise.
• It reveals good balanced efficiency on tougher instruction following benchmarks like COLLIE.
• The mannequin reveals good outcomes at MATH500 however reveals combined efficiency on MBPP which is a coding benchmark. 
• The spotlight of its benhcmarks is the resulkt on agentic software use the place its very aggressive.
• On security benchmarks the fashions nice efficiency on parameters like harmfulness, misinformation and privateness. 

How did LongCat-Flash Really Carry out?

The mannequin fell wanting expectations. With the LLM panorama advancing quickly, there may be little room for mediocrity. LongCat-Flash generated unnecessarily lengthy code, struggled with a reasoning drawback, and didn’t exhibit its agentic capabilities, producing no related outcomes.

Different Current Articles on Chinese language LLMs:

Conclusion

The mannequin reveals sturdy potential. Meituan’s first launch, LongCat-Flash, is nice however not nice. It struggles with duties that different fashions right this moment deal with with ease. Nonetheless, as a basis mannequin, it’s anticipated to enhance. That is solely the start, and with its superior structure and coaching method, it may pave the best way for a lot of stronger successors. For now, LongCat-Flash might not appear “flashy” sufficient, however within the coming months, it may emerge as a critical contender for the highest spot.

Anu Madan is an skilled in tutorial design, content material writing, and B2B advertising and marketing, with a expertise for reworking advanced concepts into impactful narratives. Together with her give attention to Generative AI, she crafts insightful, revolutionary content material that educates, evokes, and drives significant engagement.