Alibaba’s new Qwen3-235B-A22B-2507 beats Kimi-2, Claude Opus

Chinese language e-commerce large Alibaba has made waves globally within the tech and enterprise communities with its household of “Qwen” gen AI massive language fashions (LLMs), starting with the launch of the Tongyi Qianwen chatbot in April 2023, by way of the discharge of Qwen 3 in April 2025.

Why?

Nicely, not solely are its fashions highly effective and rating excessive on third-party benchmark checks for math, science, reasoning and writing duties, for essentially the most half, they’ve been launched underneath permissive open-source licensing phrases, permitting organizations and enterprises to obtain, customise, run and customarily use them for quite a lot of functions, even business. Consider them as a substitute for DeepSeek.

This week, Alibaba’s Qwen Workforce launched the newest updates to its Qwen household, and so they’re already attracting consideration from AI energy customers within the West for his or her high efficiency. In a single case, they edged out the brand new Kimi-2 mannequin from rival Chinese language AI startup Moonshot, launched in mid-July 2025.

The new Qwen3-235B-A22B-2507-Instruct mannequin — launched on AI code sharing neighborhood Hugging Face alongside a “floating point 8” or FP8 version, which we’ll cowl extra in-depth under — improves on the unique Qwen 3 in reasoning duties, factual accuracy and multilingual understanding. It additionally outperforms Claude Opus 4’s “non-thinking” model.

The brand new Qwen3 mannequin replace additionally delivers higher coding outcomes, alignment with consumer preferences and long-context dealing with, in keeping with its creators. However that’s not all…

Learn on for what it provides enterprise customers and technical decision-makers.

FP8 model lets enterprises run Qwen 3 with far much less reminiscence and compute

The “FP8” model’s 8-bit floating level compresses the mannequin’s numerical operations to make use of much less reminiscence and processing energy — with out noticeably affecting its efficiency.

In apply, this implies organizations can run a mannequin with Qwen3’s capabilities on smaller, inexpensive {hardware}, or extra effectively within the cloud. The result’s quicker response instances, decrease power prices and the power to scale deployments without having huge infrastructure.

This makes the FP8 mannequin particularly engaging for manufacturing environments with tight latency or price constraints. Groups can scale Qwen3’s capabilities to single-node GPU cases or native improvement machines, avoiding the necessity for large multi-GPU clusters. It additionally lowers the barrier to non-public fine-tuning and on-premises deployments, the place infrastructure sources are finite and whole price of possession issues.

Despite the fact that Qwen’s crew didn’t launch official calculations, comparisons to comparable FP8 quantized deployments recommend the effectivity financial savings are substantial. Right here’s a sensible illustration (up to date and corrected on 07/23/2025 at 16:04 pm ET — this piece initially included an inaccurate chart based mostly on a miscalculation. I apologize for the errors and thank readers for contacting me about them.):

Metric	BF16 / BF16-equiv construct	FP8 Quantized construct
GPU reminiscence use*	≈ 640 GB whole (8 × H100-80 GB, TP-8)	≈ 320 GB whole on the really useful 4 × H100-80 GB, TP-4    Lowest-footprint neighborhood run: ~143 GB throughout 2 × H100 with Ollama off-loading
Single-query inference pace†	~74 tokens / s (batch = 1, context = 2 okay, 8 × H20-96 GB, TP-8)	~72 tokens / s (identical settings, 4 × H20-96 GB, TP-4)
Energy / power	Full node of eight H100s attracts ~4-4.5 kW underneath load (550–600 W per card, plus host)‡	FP8 wants half the playing cards and strikes half the information; Nvidia’s Hopper FP8 case-studies report ≈ 35-40 % decrease TCO and power at comparable throughput
GPUs wanted (sensible)	8 × H100-80 GB (TP-8) or 8 × A100-80 GB for parity	4 × H100-80 GB (TP-4). 2 × H100 is feasible with aggressive off-loading, at the price of latency

*Disk footprint for the checkpoints: BF16 weights are ~500 GB; the FP8 checkpoint is “nicely over 200 GB,” so absolutely the reminiscence financial savings on GPU come principally from needing fewer playing cards, not from weights alone.

†Pace figures are from the Qwen3 official SGLang benchmarks (batch 1). Throughput scales virtually linearly with batch measurement: Baseten measured ~45 tokens/s per consumer at batch 32 and ~1.4 okay tokens/s mixture on the identical four-GPU FP8 setup.

‡No vendor provides actual wall-power numbers for Qwen, so we approximate utilizing H100 board specs and NVIDIA Hopper FP8 energy-saving information.

No extra ‘hybrid reasoning’… as an alternative, Qwen will launch separate reasoning and instruct fashions

Maybe most fascinating, Qwen introduced it’ll not be pursuing a “hybrid” reasoning method, which it launched with Qwen 3 in April. It appeared to be impressed by an method pioneered by sovereign AI collective Nous Analysis.

This allowed customers to toggle on a “reasoning” mannequin, letting the AI mannequin have interaction in its personal self-checking and producing chains-of-thought (CoT) earlier than responding.

In a means, it was designed to imitate the reasoning capabilities of highly effective proprietary fashions akin to OpenAI’s “o” collection (o1, o3, o4-mini, o4-mini-high), which additionally produce “chains-of-thought.”

Nonetheless, not like these rival fashions which at all times have interaction in such “reasoning” for each immediate, Qwen 3 can have the reasoning mode manually switched on or off with a “Considering Mode” button on the Qwen web site chatbot. Or, customers can sort “/assume” earlier than their immediate on a neighborhood or privately run mannequin inference.

The thought was to present customers management to have interaction the slower and extra token-intensive considering mode for tougher prompts and duties, and use a non-thinking mode for less complicated prompts. However once more, this put the onus on the consumer to resolve. Whereas versatile, it additionally launched design complexity and inconsistent habits in some instances.

Now As Qwen wrote on X:

“After speaking with the neighborhood and considering it by way of, we determined to cease utilizing hybrid considering mode. As an alternative, we’ll prepare Instruct and Considering fashions individually so we are able to get the highest quality attainable.”

With the 2507 replace — an instruct or non-reasoning mannequin, for now — Alibaba is not straddling each approaches in a single mannequin. As an alternative, separate mannequin variants will probably be skilled for instruction and reasoning duties, respectively.

The result’s a mannequin that adheres extra intently to consumer directions, generates extra predictable responses and, as benchmark information exhibits, improves considerably throughout a number of analysis domains.

Efficiency benchmarks and use instances

In comparison with its predecessor, the Qwen3-235B-A22B-Instruct-2507 mannequin delivers measurable enhancements:

• MMLU-Professional scores rise from 75.2 to 83.0, a notable achieve basically data efficiency.
• GPQA and SuperGPQA benchmarks enhance by 15–20 share factors, reflecting stronger factual accuracy.
• Reasoning duties akin to AIME25 and ARC-AGI present greater than double the earlier efficiency.
• Code era improves, with LiveCodeBench scores growing from 32.9 to 51.8.
• Multilingual assist expands, aided by improved protection of long-tail languages and higher alignment throughout dialects.

The mannequin maintains a mixture-of-experts (MoE) structure, activating 8 out of 128 consultants throughout inference, with a complete of 235 billion parameters — 22 billion of that are lively at any time.

As talked about, the FP8 model introduces fine-grained quantization for higher inference pace and decreased reminiscence utilization.

Enterprise-ready by design

Not like many open-source LLMs, which are sometimes launched underneath restrictive research-only licenses or require API entry for business use, Qwen3 is squarely aimed toward enterprise deployment.

Boasting a permissive Apache 2.0 license, this implies enterprises can use it freely for business functions. They might additionally:

• Deploy fashions regionally or by way of OpenAI-compatible APIs utilizing vLLM and SGLang;
• High quality-tune fashions privately utilizing LoRA or QLoRA with out exposing proprietary information;
• Log and examine all prompts and outputs on-premises for compliance and auditing;
• Scale from prototype to manufacturing utilizing dense variants (from 0.6B to 32B) or MoE checkpoints.

Alibaba’s crew additionally launched Qwen-Agent, a light-weight framework that abstracts software invocation logic for customers constructing agentic programs.

Benchmarks like TAU-Retail and BFCL-v3 recommend the instruction mannequin can competently execute multi-step resolution duties — usually the area of purpose-built brokers.

Group and trade reactions

The discharge has already been nicely acquired by AI energy customers.

Paul Couvert, AI educator and founding father of non-public LLM chatbot host Blue Shell AI, posted on X a comparability chart exhibiting Qwen3-235B-A22B-Instruct-2507 outperforming Claude Opus 4 and Kimi K2 on benchmarks like GPQA, AIME25 and Enviornment-Exhausting v2, calling it “much more highly effective than Kimi K2… and even higher than Claude Opus 4.”

AI influencer NIK (@ns123abc) commented on its fast affect: “Qwen-3-235B made Kimi K2 irrelevant after just one week, regardless of being one quarter the scale, and also you’re laughing.”

In the meantime, Jeff Boudier, head of product at Hugging Face, highlighted the deployment advantages: “Qwen silently launched an enormous enchancment to Qwen3… it tops greatest open (Kimi K2, a 4x bigger mannequin) and closed (Claude Opus 4) LLMs on benchmarks.”

He praised the supply of an FP8 checkpoint for quicker inference, 1-click deployment on Azure ML and assist for native use by way of MLX on Mac or INT4 builds from Intel.

The general tone from builders has been enthusiastic, because the mannequin’s stability of efficiency, licensing and deployability appeals to each hobbyists and professionals.

What’s subsequent for Qwen crew?

Alibaba is already laying the groundwork for future updates. A separate reasoning-focused mannequin is within the pipeline, and the Qwen roadmap factors towards more and more agentic programs able to long-horizon job planning.

Multimodal assist, as seen in Qwen2.5-Omni and Qwen-VL fashions, can also be anticipated to increase additional.

And already, rumors and rumblings have begun as Qwen crew members tease yet one more replace to their mannequin household, with their web properties revealing URL strings for a brand new Qwen3-Coder-480B-A35B-Instruct mannequin, doubtless a 480-billion parameter MoE with a token context of 1 million.

What Qwen3-235B-A22B-Instruct-2507 finally alerts is not only one other leap in benchmark efficiency, however a maturation of open fashions as viable options to proprietary programs.

The flexibleness of deployment, robust common efficiency and enterprise-friendly licensing give the mannequin a novel edge in a crowded discipline.

For groups seeking to combine superior instruction-following fashions into their AI stack — with out the restrictions of vendor lock-in or usage-based charges — Qwen3 is a severe contender.