Alibaba's small, open supply Qwen3.5-9B beats OpenAI's gpt-oss-120B and might run on normal laptops

Regardless of political turmoil within the U.S. AI sector, in China, the AI advances are persevering with apace with no hitch.

Earlier immediately, e-commerce large Alibaba's Qwen Staff of AI researchers, centered totally on growing and releasing to the world a rising household of highly effective and succesful Qwen open supply language and multimodal AI fashions, unveiled its latest batch, the Qwen3.5 Small Mannequin Collection, which consists of:

• Qwen3.5-0.8B & 2B: Two fashions, each ptimized for "tiny" and "quick" efficiency, meant for prototyping and deployment on edge gadgets the place battery life is paramount.

• Qwen3.5-4B: A robust multimodal base for light-weight brokers, natively supporting a 262,144 token context window.

• Qwen3.5-9B a compact reasoning mannequin that outperforms the 13.5x bigger U.S. rival OpenAI's open soruce gpt-oss-120B on key third-party benchmarks together with multilingual information and graduate-level reasoning

To place this into perspective, these fashions are on the order of the smallest normal goal fashions these days shipped by any lab around the globe, comparable extra to MIT offshoot LiquidAI's LFM2 sequence, which even have a number of hundred million or billion parameters, than the estimated trillion parameters (mannequin settings) reportedly used for the flagship fashions from OpenAI, Anthropic, and Google's Gemini sequence.

The weights for the fashions can be found proper now globally beneath Apache 2.0 licenses — good for enterprise and industrial use, together with customization as wanted — on Hugging Face and ModelScope.

The expertise: hybrid effectivity and native multimodality

The technical basis of the Qwen3.5 small sequence is a departure from normal Transformer architectures. Alibaba has moved towards an Environment friendly Hybrid Structure that mixes Gated Delta Networks (a type of linear consideration) with sparse Combination-of-Specialists (MoE).

This hybrid method addresses the "reminiscence wall" that sometimes limits small fashions; through the use of Gated Delta Networks, the fashions obtain greater throughput and considerably decrease latency throughout inference.

Moreover, these fashions are natively multimodal. In contrast to earlier generations that "bolted on" a imaginative and prescient encoder to a textual content mannequin, Qwen3.5 was educated utilizing early fusion on multimodal tokens. This enables the 4B and 9B fashions to exhibit a degree of visible understanding—resembling studying UI parts or counting objects in a video—that beforehand required fashions ten occasions their dimension.

Benchmarking the "small" sequence: efficiency that defies scale

Newly launched benchmark knowledge illustrates simply how aggressively these compact fashions are competing with—and sometimes exceeding—a lot bigger trade requirements. The Qwen3.5-9B and Qwen3.5-4B variants show a cross-generational leap in effectivity, significantly in multimodal and reasoning duties.

Multimodal dominance: Within the MMMU-Professional visible reasoning benchmark, Qwen3.5-9B achieved a rating of 70.1, outperforming Gemini 2.5 Flash-Lite (59.7) and even the specialised Qwen3-VL-30B-A3B (63.0).

Graduate-level reasoning: On the GPQA Diamond benchmark, the 9B mannequin reached a rating of 81.7, surpassing gpt-oss-120b (80.1), a mannequin with over ten occasions its parameter rely.

Video understanding: The sequence reveals elite efficiency in video reasoning. On the Video-MME (with subtitles) benchmark, Qwen3.5-9B scored 84.5 and the 4B scored 83.5, considerably main over Gemini 2.5 Flash-Lite (74.6).

Mathematical prowess: Within the HMMT Feb 2025 (Harvard-MIT arithmetic event) analysis, the 9B mannequin scored 83.2, whereas the 4B variant scored 74.0, proving that high-level STEM reasoning not requires large compute clusters.

Doc and multilingual information: The 9B variant leads the pack in doc recognition on OmniDocBench v1.5 with a rating of 87.7. In the meantime, it maintains a top-tier multilingual presence on MMMLU with a rating of 81.2, outperforming gpt-oss-120b (78.2).

Group reactions: "extra intelligence, much less compute"

Approaching the heels of final week's launch of an already fairly small, highly effective open supply Qwen3.5-Medium able to working on a single GPU, the announcement of the Qwen3.5-Small Fashions Collection and their even smaller footprint and processing necessities sparked rapid curiosity amongst builders centered on "local-first" AI.

"Extra intelligence, much less compute" resonated with customers in search of alternate options to cloud-based fashions.

AI and tech educator Paul Couvert of Blueshell AI captured the trade's shock relating to this effectivity leap.

"How is that this even potential?!" Couvert wrote on X. "Qwen has launched 4 new fashions and the 4B model is nearly as succesful because the earlier 80B A3B one. And the 9B is nearly as good as GPT OSS 120b whereas being 13x smaller!"

Couvert's evaluation highlights the sensible implications of those architectural positive factors:

• "They will run on any laptop computer"

• "0.8B and 2B to your cellphone"

• "Offline and open supply"

As developer Karan Kendre of Kargul Studio put it: "these fashions [can run] regionally on my M1 MacBook Air without cost."

This sentiment of "superb" accessibility is echoed throughout the developer ecosystem. One person famous {that a} 4B mannequin serving as a "sturdy multimodal base" is a "sport changer for cellular devs" who want screen-reading capabilities with out excessive CPU overhead.

Certainly, Hugging Face developer Xenova famous that the brand new Qwen3.5 Small Mannequin sequence may even run straight in a person's internet browser and carry out such subtle and beforehand higher-compute demanding operations like video evaluation.

Researchers additionally praised the discharge of Base fashions alongside the Instruct variations, noting that it supplies important assist for "real-world industrial innovation."

The discharge of Base fashions is especially valued by enterprise and analysis groups as a result of it supplies a "clean slate" that hasn't been biased by a selected set of RLHF (Reinforcement Studying from Human Suggestions) or SFT (Supervised Advantageous-Tuning) knowledge, which might typically result in "refusals" or particular conversational types which can be troublesome to undo.

Now, with the Base fashions, these fascinated with customizing the mannequin to suit particular duties and functions a neater place to begin, as they will now apply their very own instruction tuning and post-training with out having to strip away Alibaba's.

Licensing: a win for the open ecosystem

Alibaba has launched the weights and configuration information for the Qwen3.5 sequence beneath the Apache 2.0 license. This permissive license permits for industrial use, modification, and distribution with out royalty funds, eradicating the "vendor lock-in" related to proprietary APIs.

• Industrial use: Builders can combine fashions into industrial merchandise royalty-free.

• Modification: Groups can fine-tune (SFT) or apply RLHF to create specialised variations.

• Distribution: Fashions will be redistributed in local-first AI purposes like Ollama.

Contextualizing the information: why small issues a lot proper now

The discharge of the Qwen3.5 Small Collection arrives at a second of "Agentic Realignment." We’ve got moved previous easy chatbots; the purpose now’s autonomy. An autonomous agent should "suppose" (purpose), "see" (multimodality), and "act" (instrument use). Whereas doing this with trillion-parameter fashions is prohibitively costly, an area Qwen3.5-9B can carry out these loops for a fraction of the price.

By scaling Reinforcement Studying (RL) throughout million-agent environments, Alibaba has endowed these small fashions with "human-aligned judgment," permitting them to deal with multi-step goals like organizing a desktop or reverse-engineering gameplay footage into code. Whether or not it’s a 0.8B mannequin working on a smartphone or a 9B mannequin powering a coding terminal, the Qwen3.5 sequence is successfully democratizing the "agentic period."

The Qwen3.5 sequence shift from "chatbits" to "native multimodal brokers" transforms how enterprises can distribute intelligence. By shifting subtle reasoning to the "edge"—particular person gadgets and native servers—organizations can automate duties that beforehand required costly cloud APIs or high-latency processing.

Strategic enterprise purposes and concerns

The 0.8B to 9B fashions are re-engineered for effectivity, using a hybrid structure that activations solely the required elements of the community for every process.

• Visible Workflow Automation: Utilizing "pixel-level grounding," these fashions can navigate desktop or cellular UIs, fill out types, and arrange information based mostly on pure language directions.

• Advanced Doc Parsing: With scores exceeding 90% on doc understanding benchmarks, they will change separate OCR and format parsing pipelines to extract structured knowledge from various types and charts.

• Autonomous Coding & Refactoring: Enterprises can feed complete repositories (as much as 400,000 traces of code) into the 1M context window for production-ready refactors or automated debugging.

• Actual-Time Edge Evaluation: The 0.8B and 2B fashions are designed for cellular gadgets, enabling offline video summarization (as much as 60 seconds at 8 FPS) and spatial reasoning with out taxing battery life.

The desk under outlines which enterprise capabilities stand to achieve essentially the most from native, small-model deployment.

Whereas these fashions are extremely succesful, their small scale and "agentic" nature introduce particular operational "flags" that groups should monitor.

• The Hallucination Cascade: In multi-step "agentic" workflows, a small error in an early step can result in a "cascade" of failures the place the agent pursues an incorrect or nonsensical plan.

• Debugging vs. Greenfield Coding: Whereas these fashions excel at writing new "greenfield" code, they will wrestle with debugging or modifying current, advanced legacy techniques.

• Reminiscence and VRAM Calls for: Even "small" fashions (just like the 9B) require vital VRAM for high-throughput inference; the "reminiscence footprint" stays excessive as a result of the full parameter rely nonetheless occupies GPU area.

• Regulatory & Information Residency: Utilizing fashions from a China-based supplier might elevate knowledge residency questions in sure jurisdictions, although the Apache 2.0 open-weight model permits for internet hosting on "sovereign" native clouds.

Enterprises ought to prioritize "verifiable" duties—resembling coding, math, or instruction following—the place the output will be routinely checked towards predefined guidelines to forestall "reward hacking" or silent failures.