Latest developments on AI fashions have been particularly targeted on one job – coding brokers. Following the road, Alibaba’s Qwen is now out with a brand new mannequin, one which guarantees industry-leading outcomes, all whereas operating domestically. This basically makes it an open-weight language mannequin designed, within the phrases of the Qwen crew, “particularly for coding brokers and native growth.” As for the moniker, Alibaba’s AI division has conveniently termed the brand new mannequin Qwen3-Coder-Subsequent.
So what makes the brand new mannequin completely different? Sharing a touch at this in a weblog, the crew at Qwen particularly mentions that the Qwen3-Coder-Subsequent has been “agentically skilled at scale on large-scale executable job synthesis, setting interplay, and reinforcement studying.” Due to this coaching, the Qwen3-Coder-Subsequent is claimed to come back with “robust coding and agentic capabilities,” all whereas incurring a “considerably decrease inference prices.”
Simply how a lot enchancment is that this within the real-world use circumstances? Let’s discover out on this article.
What’s Qwen3-Coder-Subsequent?
As talked about, Qwen3-Coder-Subsequent is an open-weight language mannequin, which suggests its skilled parameters (weights) are launched publicly, and anybody can obtain, run, and fine-tune it domestically (learn extra about open-weight fashions right here). At its core, Qwen3-Coder-Subsequent is constructed to behave much less like a chatbot and extra like an actual software program agent. As a substitute of simply predicting the following line of code, it’s skilled to grasp your targets, work together with execution environments, and iterate towards working options. Whereas it might not sound a lot to you, this distinction issues quite a bit.
A significant spotlight with the mannequin is that it’s purpose-built for agentic workflows. This implies Qwen3-Coder-Subsequent can plan multi-step duties, purpose over lengthy recordsdata, run code, observe outputs, and regulate its method. To think about it, that is precisely how human builders truly work.
Yet one more key level is deployment. Qwen3-Coder-Subsequent is designed to run domestically, giving builders full management over their setting, knowledge, and workflows. That makes it particularly enticing for groups engaged on proprietary codebases, inner tooling, or offline setups, with out sacrificing functionality.
Briefly, this isn’t only a “write me a perform” mannequin. Qwen3-Coder-Subsequent is Qwen’s try to show AI right into a sensible, domestically runnable coding agent. One that may assume, take a look at, and enhance its personal output.
Qwen3-Coder-Subsequent Structure
In its weblog, the Qwen crew specifies that the Qwen3-Coder-Subsequent has been “constructed on high of Qwen3-Subsequent-80B-A3B-Base.” This implies it inherits the identical hybrid consideration and MoE (Combination of Specialists) structure that powers Qwen’s newest technology of huge fashions.
This hybrid consideration design is among the many finest strategies for balancing long-context understanding with environment friendly computation. Which implies, as a substitute of making use of full consideration in every single place, it selectively allocates consideration the place it issues most. That is particularly necessary for coding duties, the place the mannequin should purpose throughout giant recordsdata, dependencies, and execution logs with out blowing up reminiscence or latency.
The MoE setup additional sharpens this effectivity. Moderately than activating all the mannequin for each token, Qwen3-Coder-Subsequent dynamically routes duties to a small subset of specialised “specialists.” In observe, this implies you get the advantages of a really giant mannequin, however you solely pay the compute value of a a lot smaller one throughout inference.
Now let’s revert to the 2 core competencies of Qwen3-Coder-Subsequent. The mannequin can assist agentic coding workflows at scale, whereas nonetheless being sensible sufficient to run domestically. It’s clear that that is potential solely by way of these architectural decisions.
Now that we all know how its constructed, let’s try its efficiency on benchmark scores.
Qwen3-Coder-Subsequent Benchmark Efficiency
Primarily based on the official benchmarks (learn extra about AI benchmarks right here) shared by the Qwen crew, right here is how Qwen3-Coder-Subsequent performs throughout broadly used coding agent evaluations:
- SWE-Bench Verified (with SWE-Agent): 70.6% success price
- SWE-Bench Multilingual (with SWE-Agent): 62.8% success price
- SWE-Bench Professional (with SWE-Agent): 44.3% success price
- Terminal-Bench 2.0 (with Terminus-2 JSON): 36.2% success price
- Aider benchmark: 66.2% success price
What These Scores Inform Us
The standout efficiency on SWE-Bench Verified proves that Qwen3-Coder-Subsequent is extremely efficient at real-world software program upkeep duties, particularly these involving bug fixing and repository-level reasoning. This benchmark intently mirrors what builders face in manufacturing codebases, making this outcome notably significant.
Its robust displaying on SWE-Bench Multilingual highlights one other key energy: the mannequin isn’t restricted to English-only code contexts. It might purpose throughout multilingual repositories, feedback, and documentation, and never drop its consistency. That is an more and more necessary requirement for international growth groups.
The SWE-Bench Professional rating additional reinforces that this mannequin is constructed for agentic depth, whereas Terminal-Bench 2.0 outcomes point out dependable command-line reasoning and structured software interplay. Lastly, the robust efficiency on Aider, a benchmark targeted on AI-assisted coding workflows, exhibits that Qwen3-Coder-Subsequent integrates nicely into actual developer tooling.
If we had been to summarise this benchmark efficiency, it’s clear that the brand new Qwen mannequin is optimized for sensible coding brokers. Its efficiency persistently displays the flexibility to plan, act, observe, and iterate. And that is precisely what fashionable AI-powered growth workflows demand.
Qwen3-Coder-Subsequent: Learn how to Entry
There are 3 methods you may entry the brand new Qwen3-Coder-Subsequent, based mostly on the platform you need it on –
HuggingFace – https://huggingface.co/collections/Qwen/qwen3-coder-next
Kaggle – https://www.kaggle.com/fashions/qwen-lm/qwen3-coder-next
ModelScope – https://modelscope.cn/collections/Qwen/Qwen3-Coder-Subsequent
Palms-on with Qwen3-Coder-Subsequent
To check out the prowess of the most recent Qwen mannequin, we put it by way of some real-world exams. We shared the next prompts with it to take a look at the outputs, and here’s what we bought.
1. A Sport of Snake
Immediate:
Create a easy snake recreation. embrace bonus factors in between that develop the snake by 3 factors as a substitute of 1
Output:
Snake Sport - Bonus Mode
Regular (+1)
Bonus (+3)
Use Arrow Keys to maneuver
Qwen3-Coder-Subsequent was capable of generate this complete code inside a cut up second. And having examined it out, I’ve discovered it to be working completely, even with the particular instruction of bonus factors that develop the snake disproportionately. Go forward, attempt a immediate like this and you’ll be amazed on the pace and accuracy of the mannequin.
2. HTML Code for Easy Animation
Immediate:
Give me an HTML code for a circle rotating inside a sq., which is inside a triangle. The ball and triangle are rotating in a single course and sq. within the reverse.
Output:
Rotating Geometric Shapes
But once more, Qwen3-Coder-Subsequent got here up with a code for the duty at hand at lightning pace. It precisely depicts the shapes and their rotation. The one little change I’d’ve needed right here was for the shapes to be fully contained in the outer one, versus the half-in, half-out association as proven right here. Aside from that, the AI mannequin did it brilliantly nicely.
3. Primary HTML Web site
Immediate:
Construct a fundamental HTML web site for a tourism firm, itemizing packages for the favored vacationer locations in India. Add one other class displaying journey actions as packages. Preserve a header with a emblem and the everyday menu of an internet site. Hold the color theme blue and white.
Output:
India Travels | Discover the Pink & Blue Nation
From the majestic Himalayas to the serene backwaters of Kerala, expertise the colours, tradition, and chaos of the subcontinent.
Discover Packages
Northern Circuit
The Golden Triangle
Delhi, Agra, and Jaipur. Discover Mughal structure, fortresses, and the colourful bazaars of Rajasthan.
7 Days
4 Cities
Resort Included
$350
West Coast
Goa Seashore & Celebration Tour
Loosen up on sun-kissed seashores, expertise vibrant night time life, and luxuriate in water sports activities in South Goa and North Goa.
5 Days
Seashore Resorts
Events Included
$280
Southern Serenity
Kerala Backwaters & Wildlife
Cruise on houseboats in Alleppey, go to spice plantations, and spot tigers in Periyar Wildlife Sanctuary.
6 Days
Houseboat Keep
Spice Excursions
$420
Bungee & River
Rishikesh Journey Camp
Bungee leaping, river rafting, and mountain climbing within the yoga capital of the world, nestled within the Himalayas.
2 Days
Camp Keep
Skilled Gear
$120
Excessive Altitude
Ladakh Bike Safari
Experience by way of the world's highest motorable roads, cross Shinga La cross, and see the pristine Pangong Lake.
8 Days
Self-Drive Bike
Information Included
$600
Desert Safari
Thar Desert Camp (Jaisalmer)
Camel driving at sundown, conventional Rajasthani dinner below the celebrities, and sleeping in luxurious tented camps.
1 Day
Camel Experience
Dinner & Dance
$80
That is hands-down the most effective AI-coded web sites I’ve seen to this point. It tends to have the whole lot in place, with an correct depiction of all directions in addition to the weather that I didn’t actually specify. The color scheme is on level. The spacing and sections are nicely laid out, and the textual content is simply excellent to learn. That is the sort of output that may take you straight from a immediate to going dwell, so kudos to Qwen3-Coder-Subsequent for the tremendous fast and near-perfect output.
Conclusion
At first look, Qwen3-Coder-Subsequent makes some large guarantees – coding brokers, strong reasoning, and the flexibility to run domestically at a minimal inference value. However after you have a take a look at its benchmark efficiency, you understand these aren’t simply phrases thrown within the air. These are backed by some very actual efficiency metrics.
And that’s while you really feel that the brand new Qwen mannequin is already sensible sufficient for real-world growth workflows. That stated, that is clearly not the top of the street. Whereas the mannequin performs competitively even towards a lot bigger open-source methods, the Qwen crew is upfront in regards to the room for enchancment. And that honesty issues.
Wanting forward, Qwen’s focus is firmly on strengthening agent expertise: higher reasoning, smarter software use, broader job protection. It even guarantees fast updates based mostly on “how folks use it.” Now that’s the sort of dedication you want from the makers of a fantastic AI mannequin. And if this trajectory holds, it received’t be a marvel when Qwen3-Coder-Subsequent will evolve from a robust native coding assistant right into a genuinely autonomous software program agent.
Login to proceed studying and luxuriate in expert-curated content material.
