---
title: DeepSeek-R1-Distill-Qwen-1.5B Model
url: https://www.emergentmind.com/topics/deepseek-r1-distill-qwen-1-5b
type: topic
---

# DeepSeek-R1-Distill-Qwen-1.5B Model

DeepSeek-R1-Distill-Qwen-1.5B is a 1.5-billion-parameter dense language model constructed via supervised distillation from the DeepSeek-R1 teacher, itself a reinforcement-learning-optimized reasoning LLM. This model, built on the Qwen2.5 backbone, is engineered to inherit advanced chain-of-thought reasoning abilities suitable for competitive mathematical, logical, and code tasks, while being sufficiently compact for resource-limited deployment. Its training, evaluation, and subsequent optimization illustrate core trends in the RL-for-reasoning LLM paradigm, efficiency-oriented distillation, and fine-grained reward shaping.

## 1. Origin and Distillation Methodology

DeepSeek-R1-Distill-Qwen-1.5B is produced by distilling the reasoning behaviors and chain-of-thought outputs of the DeepSeek-R1 model into the Qwen2.5-1.5B backbone through supervised fine-tuning (SFT) [2501.12948]. DeepSeek-R1 itself emerges from a pipeline consisting of a cold-start phase (high-quality, long CoT SFT), RL (predominantly GRPO-based), and further SFT, producing reasoning chains delimited by tags such as `<think>` and `<answer>`. 

The distillation leverages approximately 800,000 high-quality reasoning traces generated from DeepSeek-R1, sourced from challenging mathematical, logical, and programmatic tasks. No additional RL is performed during distillation at the 1.5B scale; SFT alone suffices to transfer chain-of-thought paradigms and reasoning structure. The Qwen series is favored as the student backbone for its alignment properties and, specifically in Qwen2.5, robust base reasoning skills.

## 2. Reasoning Capability and Performance Scaling

The primary contribution of DeepSeek-R1-Distill-Qwen-1.5B is the reliable transfer of complex multi-step reasoning to a parameter-efficient model. On logical reasoning benchmarks, such as AIME 2024 and MATH-500, the distilled 1.5B variant demonstrates substantial gains over math-focused or instruction-tuned baselines and even surpasses larger non-reasoning LLMs in discriminator-oriented tasks [2502.11164][2505.03786]. For instance, distillation increased scores on general tasks for Qwen2.5-Math-1.5B by 178.74%.

Empirical results in several works show that while scaling the model parameter count (e.g., from 1.5B up to 32B/70B) yields monotonic accuracy improvements per established scaling laws, the benefits of reasoning-driven SFT disproportionately elevate smaller models, particularly when the training data is curated for difficulty and diversity [2502.11164]. Notably, the DeepSeek-R1-Distill-Qwen-1.5B model can achieve better discrimination accuracy than much larger models, thus inverting scale-driven performance expectations in evaluation-heavy or discriminator roles.

Performance across tasks:

| Model                         | AIME2024 | MATH-500 | LiveCodeBench | GPQA-Diamond |
|-------------------------------|----------|----------|---------------|--------------|
| DeepSeek-R1-Distill-Qwen-1.5B | 70.9     | 95.8     | 57.0          | —            |
| Qwen3-235B-A22B-distill       | 79.4     | 93.9     | 59.6          | —            |
| AM-Thinking-v1-distill        | 84.3     | 98.4     | 65.9          | —            |

Relative to its scale, DeepSeek-R1-Distill-Qwen-1.5B is competitive in logical reasoning and mathematical correctness.

## 3. Training Data Curation and Quality

The dataset for distillation includes high-quality, diversity-enhanced reasoning traces with rigorous verification: mathematical answers are checked via exact or epsilon-bound matching, code responses are validated using sandboxed test cases, and open QA traces are filtered through scoring reward models [2503.19633]. This data is structured with explicit `<think>` and `<answer>` blocks, and often includes metadata and LaTeX-marked derivations, crucial for inducting stepwise, verifiable reasoning.

Dataset diversity across domains (math, code, science, general chat) and semantic deduplication further mitigate data contamination and n-gram redundancy. However, comparative studies show that, although DeepSeek-R1-derived traces are valuable, datasets distilled from more length-diverse and lower-perplexity teachers (e.g., AM-Thinking-v1) yield superior adaptation and benchmark performance [2505.14464].

## 4. Task Specialization, Efficiency, and Reward Strategies

DeepSeek-R1-Distill-Qwen-1.5B is particularly effective when used as a discriminator in planning frameworks: its structured chain-of-thought traces can be parsed for soft-scoring candidate outputs (e.g., in text-to-SQL ranking), giving it pronounced advantages even over upscaled general LLMs [2505.03786]. For generation tasks, especially those requiring fluent or concise open-ended outputs, model performance is more modest, and further refinement may be necessary.

To address reasoning efficiency, several RL-based reward shaping techniques (LASER-D, HAPO, AutoThink, Flexible Realignment) have been applied post-distillation to reduce extraneous tokens and over-thinking:

- **LASER-D**: Dynamically adjusts target reasoning lengths by difficulty bucket, penalizing lengthy outputs for easy tasks and rewarding concise correct traces. Achieves up to +6.1 accuracy on AIME2024 while reducing tokens by 63% [2505.15612].
- **HAPO**: Tracks minimal correct response lengths in training history, using a length reward combined with correctness so that new solutions are shorter and errors are not over-penalized. Delivers 49% response compression at <2% accuracy loss [2505.11225].
- **AutoThink**: Employs a simple “…“ prompt to stochastically invoke explicit reasoning only as needed, with multi-stage RL to learn difficulty-aware switching between fast (no CoT) and slow (CoT) modes, yielding a 6.4% relative accuracy gain and 52% reduction in token usage [2505.10832].
- **Flexible Realignment**: Combines training-time distillation and inference-time adapters to control alignment between the reference and efficiently-aligned models via logit fusion, achieving up to 54.63% token reduction [2506.12704].

These frameworks support adaptation across deployment scenarios—concise reasoning for lightweight inference or verbose step-by-step analysis for complex queries.

## 5. Safety Considerations and Alignment in Multilingual Contexts

Distillation from DeepSeek-R1 into smaller backbones can degrade safety alignment, particularly in non-English (e.g., Chinese) settings, as measured by risk content identification and refusal tasks. Comparative evaluation using the CHiSafetyBench indicates accuracy declines up to 30% in discrimination-related metrics after distillation [2503.16529]. 

To remediate this, post-distillation safety tuning on a dedicated 50,000-sample corpus containing safety-critical instructions and chain-of-thought examples appreciably recovers the model’s risk aversion, refusal accuracy, and reduces harm rates, introducing minimal (<3%) reasoning performance loss. Safety-enhanced DeepSeek-R1-Distill-Qwen variants are open-sourced to encourage best practices in safe reasoning LLM deployment.

## 6. Evaluation Protocols and Reproducibility Challenges

Benchmarking DeepSeek-R1-Distill-Qwen-1.5B (and R1-style LLMs) is highly sensitive to seemingly minor changes in experimental setup: random seed, data versioning (e.g., image inclusion), prompt construction, and parallelism settings can each sway scores by several percentage points [2506.04734]. 

Best practices advocated include:

- Fixing seeds and reporting results over sufficient (N ≥ 64) repeated runs.
- Reporting confidence intervals for mean scores, using
  $$
  \mu \pm z_{\alpha/2} \cdot (s/\sqrt{N})
  $$
  to quantify error margins.
- Full documentation of dataset versions, prompt ordering, and implementation choices.

This rigor is paramount given that performance spreads due to noise can sometimes rival true model differences, particularly for distillation-optimized models with higher evaluation variance.

## 7. Practical Deployment and Comparative Limitations

The core strengths of DeepSeek-R1-Distill-Qwen-1.5B are its parameter efficiency, cost-effective reasoning performance, and adaptability to domain-specific discrimination roles. However, its relative ranking versus baseline instruction-tuned models is task-dependent:

- **Logical reasoning and discrimination**: Outscores larger non-reasoning LLMs (e.g., +87% F1 and +3.7% execution accuracy over CodeLlama-13B in text-to-SQL) [2505.03786].
- **Generation tasks**: May underperform, especially if brevity or open-ended creativity is paramount; further fine-tuning leveraging more length-diverse, higher-quality reasoning traces (AM-Thinking-v1-derived) can close the gap [2505.14464].
- **Schema-constrained tasks**: ThinkJSON RL pipelines combining GRPO with format and structure-aware rewards enable robust schema adherence, at some accuracy cost relative to full-scale models [2502.14905].
- **Safety**: Post-distillation safety tuning is essential, especially for regulated or multilingual deployments.

Empirical results show that with advanced reward shaping and careful data curation, the DeepSeek-R1-Distill-Qwen-1.5B can serve both as an efficient solution for large-scale deployment and as a modular component in more complex planning or agentic frameworks.

## References

- [2501.12948] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- [2502.11164] Quantifying the Capability Boundary of DeepSeek Models: An Application-Driven Performance Analysis
- [2503.04548] An Empirical Study on Eliciting and Improving R1-like Reasoning Models
- [2503.19633] 1.4 Million Open-Source Distilled Reasoning Dataset to Empower Large Language Model Training
- [2503.16529] Safety Evaluation and Enhancement of DeepSeek Models in Chinese Contexts
- [2505.03786] When Reasoning Beats Scale: A 1.5B Reasoning Model Outranks 13B LLMs as Discriminator
- [2505.14464] Not All Correct Answers Are Equal: Why Your Distillation Source Matters
- [2505.15612] Learn to Reason Efficiently with Adaptive Length-based Reward Shaping
- [2505.11225] HAPO: Training Language Models to Reason Concisely via History-Aware Policy Optimization
- [2505.10832] Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL
- [2504.10449] M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models
- [2506.04734] Evaluation is All You Need: Strategic Overclaiming of LLM Reasoning Capabilities Through Evaluation Design
- [2506.12704] Flexible Realignment of Language Models
- [2502.14905] Think Inside the JSON: Reinforcement Strategy for Strict LLM Schema Adherence

This suite of works reflects the central position of DeepSeek-R1-Distill-Qwen-1.5B as a baseline for efficient open-source reasoning-oriented LLMs and a substrate for further research in reward design, safety, and application-specific customization.

Source: https://www.emergentmind.com/topics/deepseek-r1-distill-qwen-1-5b