Alibaba’s small, open source Qwen3.5-9B beats OpenAI’s gpt-oss-120B and can run on standard laptops

Despite political turmoil in the U.S. AI sector, in China, the AI advances are continuing apace without a hitch.

Earlier today, e-commerce giant Alibaba’s Qwen Team of AI researchers, focused primarily on developing and releasing to the world a growing family of powerful and capable Qwen open source language and multimodal AI models, unveiled its newest batch, the Qwen3.5 Small Model Series, which consists of:

• Qwen3.5-0.8B & 2B: Two models, both ptimized for “tiny” and “fast” performance, intended for prototyping and deployment on edge devices where battery life is paramount.

• Qwen3.5-4B: A strong multimodal base for lightweight agents, natively supporting a 262,144 token context window.

• Qwen3.5-9B a compact reasoning model that outperforms the 13.5x larger U.S. rival OpenAI’s open soruce gpt-oss-120B on key third-party benchmarks including multilingual knowledge and graduate-level reasoning

To put this into perspective, these models are on the order of the smallest general purpose models lately shipped by any lab around the world, comparable more to MIT offshoot LiquidAI’s LFM2 series, which also have several hundred million or billion parameters, than the estimated trillion parameters (model settings) reportedly used for the flagship models from OpenAI, Anthropic, and Google’s Gemini series.

The weights for the models are available right now globally under Apache 2.0 licenses — perfect for enterprise and commercial use, including customization as needed — on Hugging Face and ModelScope.

The technology: hybrid efficiency and native multimodality

The technical foundation of the Qwen3.5 small series is a departure from standard Transformer architectures. Alibaba has moved toward an Efficient Hybrid Architecture that combines Gated Delta Networks (a form of linear attention) with sparse Mixture-of-Experts (MoE).

This hybrid approach addresses the “memory wall” that typically limits small models; by using Gated Delta Networks, the models achieve higher throughput and significantly lower latency during inference.

Furthermore, these models are natively multimodal. Unlike previous generations that “bolted on” a vision encoder to a text model, Qwen3.5 was trained using early fusion on multimodal tokens. This allows the 4B and 9B models to exhibit a level of visual understanding—such as reading UI elements or counting objects in a video—that previously required models ten times their size.

Benchmarking the “small” series: performance that defies scale

Newly released benchmark data illustrates just how aggressively these compact models are competing with—and often exceeding—much larger industry standards. The Qwen3.5-9B and Qwen3.5-4B variants demonstrate a cross-generational leap in efficiency, particularly in multimodal and reasoning tasks.

Multimodal dominance: In the MMMU-Pro visual reasoning benchmark, Qwen3.5-9B achieved a score of 70.1, outperforming Gemini 2.5 Flash-Lite (59.7) and even the specialized Qwen3-VL-30B-A3B (63.0).

Graduate-level reasoning: On the GPQA Diamond benchmark, the 9B model reached a score of 81.7, surpassing gpt-oss-120b (80.1), a model with over ten times its parameter count.

Video understanding: The series shows elite performance in video reasoning. On the Video-MME (with subtitles) benchmark, Qwen3.5-9B scored 84.5 and the 4B scored 83.5, significantly leading over Gemini 2.5 Flash-Lite (74.6).

Mathematical prowess: In the HMMT Feb 2025 (Harvard-MIT mathematics tournament) evaluation, the 9B model scored 83.2, while the 4B variant scored 74.0, proving that high-level STEM reasoning no longer requires massive compute clusters.

Document and multilingual knowledge: The 9B variant leads the pack in document recognition on OmniDocBench v1.5 with a score of 87.7. Meanwhile, it maintains a top-tier multilingual presence on MMMLU with a score of 81.2, outperforming gpt-oss-120b (78.2).

Community reactions: “more intelligence, less compute”

Coming on the heels of last week’s release of an already pretty small, powerful open source Qwen3.5-Medium capable of running on a single GPU, the announcement of the Qwen3.5-Small Models Series and their even smaller footprint and processing requirements sparked immediate interest among developers focused on “local-first” AI.

“More intelligence, less compute” resonated with users seeking alternatives to cloud-based models.

AI and tech educator Paul Couvert of Blueshell AI captured the industry’s shock regarding this efficiency leap.

“How is this even possible?!” Couvert wrote on X. “Qwen has released 4 new models and the 4B version is almost as capable as the previous 80B A3B one. And the 9B is as good as GPT OSS 120b while being 13x smaller!”

Couvert’s analysis highlights the practical implications of these architectural gains:

• “They can run on any laptop”

• “0.8B and 2B for your phone”

• “Offline and open source”

As developer Karan Kendre of Kargul Studio put it: “these models [can run] locally on my M1 MacBook Air for free.”

This sentiment of “amazing” accessibility is echoed across the developer ecosystem. One user noted that a 4B model serving as a “strong multimodal base” is a “game changer for mobile devs” who need screen-reading capabilities without high CPU overhead.

Indeed, Hugging Face developer Xenova noted that the new Qwen3.5 Small Model series can even run directly in a user’s web browser and perform such sophisticated and previously higher-compute demanding operations like video analysis.

Researchers also praised the release of Base models alongside the Instruct versions, noting that it provides essential support for “real-world industrial innovation.”

The release of Base models is particularly valued by enterprise and research teams because it provides a “blank slate” that hasn’t been biased by a specific set of RLHF (Reinforcement Learning from Human Feedback) or SFT (Supervised Fine-Tuning) data, which can often lead to “refusals” or specific conversational styles that are difficult to undo.

Now, with the Base models, those interested in customizing the model to fit specific tasks and purposes an easier starting point, as they can now apply their own instruction tuning and post-training without having to strip away Alibaba’s.

Licensing: a win for the open ecosystem

Alibaba has released the weights and configuration files for the Qwen3.5 series under the Apache 2.0 license. This permissive license allows for commercial use, modification, and distribution without royalty payments, removing the “vendor lock-in” associated with proprietary APIs.

• Commercial use: Developers can integrate models into commercial products royalty-free.

• Modification: Teams can fine-tune (SFT) or apply RLHF to create specialized versions.

• Distribution: Models can be redistributed in local-first AI applications like Ollama.

Contextualizing the news: why small matters so much right now

The release of the Qwen3.5 Small Series arrives at a moment of “Agentic Realignment.” We have moved past simple chatbots; the goal now is autonomy. An autonomous agent must “think” (reason), “see” (multimodality), and “act” (tool use). While doing this with trillion-parameter models is prohibitively expensive, a local Qwen3.5-9B can perform these loops for a fraction of the cost.

By scaling Reinforcement Learning (RL) across million-agent environments, Alibaba has endowed these small models with “human-aligned judgment,” allowing them to handle multi-step objectives like organizing a desktop or reverse-engineering gameplay footage into code. Whether it is a 0.8B model running on a smartphone or a 9B model powering a coding terminal, the Qwen3.5 series is effectively democratizing the “agentic era.”

The Qwen3.5 series shift from “chatbits” to “native multimodal agents” transforms how enterprises can distribute intelligence. By moving sophisticated reasoning to the “edge”—individual devices and local servers—organizations can automate tasks that previously required expensive cloud APIs or high-latency processing.

Strategic enterprise applications and considerations

The 0.8B to 9B models are re-engineered for efficiency, utilizing a hybrid architecture that activations only the necessary parts of the network for each task.

• Visual Workflow Automation: Using “pixel-level grounding,” these models can navigate desktop or mobile UIs, fill out forms, and organize files based on natural language instructions.

• Complex Document Parsing: With scores exceeding 90% on document understanding benchmarks, they can replace separate OCR and layout parsing pipelines to extract structured data from diverse forms and charts.

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

• Real-Time Edge Analysis: The 0.8B and 2B models are designed for mobile devices, enabling offline video summarization (up to 60 seconds at 8 FPS) and spatial reasoning without taxing battery life.

The table below outlines which enterprise functions stand to gain the most from local, small-model deployment.

While these models are highly capable, their small scale and “agentic” nature introduce specific operational “flags” that teams must monitor.

• The Hallucination Cascade: In multi-step “agentic” workflows, a small error in an early step can lead to a “cascade” of failures where the agent pursues an incorrect or nonsensical plan.

• Debugging vs. Greenfield Coding: While these models excel at writing new “greenfield” code, they can struggle with debugging or modifying existing, complex legacy systems.

• Memory and VRAM Demands: Even “small” models (like the 9B) require significant VRAM for high-throughput inference; the “memory footprint” remains high because the total parameter count still occupies GPU space.

• Regulatory & Data Residency: Using models from a China-based provider may raise data residency questions in certain jurisdictions, though the Apache 2.0 open-weight version allows for hosting on “sovereign” local clouds.

Enterprises should prioritize “verifiable” tasks—such as coding, math, or instruction following—where the output can be automatically checked against predefined rules to prevent “reward hacking” or silent failures.