QwQ-32B large model of the Ali Open Source: Small Parameters pry large performance, training cost 1/10 of DeepSeek r1

A lot of people are firing a "general intelligence" Manus today, but I think the new QwQ-32B model that Ali has just opened is even more worth searching for.

I. Core breakthroughs and technological innovation in QwQ-32B

Aliyun released and opened a new generation of reasoning models, QwQ-32B, on March 6, 2025, with a view to:32 billion parametersIt's a very simple scale, achieved through enhanced learning techniquesDeepSeek-R1 for 671 billion parametersEquivalent performance, even in certain areas. The following are its core highlights:

Parameter efficiency revolution
The QwQ-32B parameter is only DeepSeek-R11/21, but throughTwo-stage intensive learning (RL)The feasibility of the “Small Model + Enhanced Learning” technology route was validated by training, which was performed on or beyond the latter in tasks such as mathematical reasoning (AIME24 assessment) and code generation (LiveCodeBench). This design breaks the path of traditional reliance on the stacking of parameters and marks the shift of the AI model from “scale competition” to “efficiency optimization”.
Agent's capacity for integration and dynamic reasoning
For the first time, the model was integrated deep in the reasoning model.Smart Agency Capability, support tools for mobilization, critical thinking and dynamic reasoning adjustments driven by environmental feedback. For example, in addressing multi-step mathematical issues, the model automatically calls the computing tool to validate intermediate results and optimizes the final answer through feedback cycles. This capacity makes it more flexible and reliable in the handling of complex mandates.
Low-cost deployment and inclusive applications
QwQ-32B supportsConsumer scale graphic cards (e.g. NVIDIA RTX 4070)Locally deployed, only DeepSeek-R1 deduction cost.1/10(The single output cost is approximately US$ 0.25), which significantly lowers the threshold for enterprise applications. This feature is particularly relevant for data-sensitive scenarios such as medical, financial, and peripheral computing devices.
Technological breakthroughs for enhanced learning
The Ali team rejected the traditional reward model and adopted it.Phased RL training strategy:
- Phase I: For mathematical and programming tasks, the correctness of the answer is verified directly via the code and training efficiency is enhanced.
- Phase II: Introduction of generic incentive models and rule-certifiers to expand the capabilities of models in generic missions while maintaining mathematical and programming performance. The study showed that RL training rounds increased the performance of the constant upgrade model, which validated the method.

Integrated comparison of mainstream models

Compare QwQ-32B from multiple dimensions with other mainstream models (DeepSeek-R1, OpenAI o1-Mini, Meta Llama3-400B):

Dimensions	QwQ-32B	DeepSeek-R1	OpenAI o1-mini	Meta Llama3-400B
Arguments	$32 billion.	671 billion (activated 37 billion)	It's about a hundred billion.	40 billion.
Training methods	Two-stage intensive learning (mathematics/programming generic)	Mixed expert model (MoE)	Supervision of learning	Self-return pre-training + command fine-tuning
Critical capabilities	Mathematic reasoning, code generation is the same as R1.	It's one of the best.	More common tasks, weaker math/code	Multilingual support for excellence, moderate reasoning
Deployment costs	Consumer-grade graphic cards at a cost of $0.25/kilotoken	Profession, cost $2.5/kilotoken	Cloud API call, higher cost.	High-end GPU cluster required
Open source ecology	Apache 2.0 protocol, full platform open source	Partial open source	Closed	Open-source communities are active, with fewer derivative models
Agent capacity	Support tool calls, dynamic reasoning adjustments	No specified integration	Limited Tool Support	None

Performance test bright spots

Mathematical reasoning (AIME24): QwQ-32B is the same as DeepSeek-R1 and significantly superior to o1-mini.
Code Generation (LiveCodeBench): QwQ-32B is comparable to R1, exceeding the same size distillation model by more than 50%.
LifeBech: QwQ-32B exceeds R1 (70) by 72.5 points, close to o3-mini (75) at a cost of only 1/20.

Detailed steps to deploy QwQ-32B using olama

olama is a local tool to run LLM. The following are steps to install and deploy QwQ-32B:

Download
- Visits llama of social website, download the executable according to the operating system.
Install olama:
- Runs the downloadable executable to complete installation.
Start olama server:
- Open terminal, run commands ollama serveI don't know.
Draw QwQ-32B models:
- Use command ollama pull QwQ:32b Download the model.

Run models:
- Use ollama run QwQ:32b --input "你的提示" Interact with models.

Open source ecology and community impacts

All Open Source and Developer Support
QwQ-32B Open Source of the Apache 2.0 protocol, covering platforms such as Hugging Face, ModelScope, Magic Community, etc.Local deployment tool chain (e.g. Ollama)andCloud API Service (Platform)I don't know. Developers can quickly fine-tune models, evaluate and containerize deployments (e.g. through the Aliyun PAI platform).
The world's largest open source model community
Since 2023, more than 200 models have been opened in Ali, covering all-modular forms of text, vision, voice, etc., ranging from 0.5B to 110B. The number of Qwen series derivative models has exceeded 100,000, moving beyond the Llama series into the largest open source model cluster globally. For example, the visual generation model “one-size-fits-all” and the code model “peripheric code” have been widely integrated into the industrial landscape.
Community response and business value
On the same day, the price of the Ali stock rose by 7 per cent a day, reflecting the confidence of the capital market in technology landings. The open source strategy promotes the ecological prosperity of developers, with rapid drops of QwQ-32B in the Hugging Face community, with derivative applications covering areas such as education, finance and health.

IV. Sector implications and future prospects

Technical paradigm shift
The success of QwQ-32B is evidence that AI R & D is moving from a “mixed miracle” to an “exact wisdom”, that parameter size is no longer the only indicator, and that optimization of algorithms and training methods is the new focus. For example, the revival of enhanced learning provides new ideas for the AGI path.
The exploration of intelligence and long-term reasoning
Ali plans to further integrate intelligence capabilities and exploreLong-term reasoningFor example, complex planning issues are addressed in multiple rounds of environments. This will facilitate the evolution of AI from “tools” to “collaborative partners”.
Open source competition and inclusive trends
The light quantification and low-cost deployment capacity of the model accelerates the technological inclusion of AI, and SMEs can build up a customized programme based on QwQ-32B quickly to break the mega-calculations monopoly. For example, locally deployed models in the medical field can handle sensitive patient data without having to rely on clouds.

Concluding remarks

The release of QwQ-32B is not just a technological breakthrough, but a comprehensive innovation in open-source ecological and business models. It achieves a leap forward in the performance of small-parameter models by enhancing the integration of learning and intelligence bodies; and re-engineered the competitive pattern of AI applications through the full-source and low-cost deployments. In the future, the model will be a key milestone in moving the AGI process forward with the landing of the $380 billion AI infrastructure input.