Loading page
DeepSeek
DeepSeek R1 Distill Qwen 1.5B
DeepSeek R1 Distill Qwen 1.5B is a 1.5-billion-parameter reasoning-focused chat model from DeepSeek, created by distilling DeepSeek-R1 outputs into a Qwen base model.
Input from
$0.180 / 1M tokens
across 1 provider
API Pricing
| Provider | Input / 1M | Output / 1M |
|---|---|---|
| $0.180 | $0.180 |
Prices updated daily. Last check: Sep 5, 2026
DeepSeek R1 Distill Qwen 1.5B pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
1.0 / 100
Math
22.0 / 100
Reasoning & Knowledge
- MMLU-Pro26.9%
- GPQA Diamond9.8%
- Humanity's Last Exam3.1%
Coding
- LiveCodeBench7.0%
Math
- AIME 202522.0%
- AIME17.7%
- MATH-50068.7%
Instruction & Long Context
- IFBench13.2%
- Long-Context Reasoning1.7%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- DeepSeek
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- Distilled from DeepSeek-R1 reasoning traces, so it produces explicit step-by-step chains of thought rather than direct answers only
- At approximately 1.5B parameters it fits in modest GPU memory and can run on consumer hardware or CPUs with quantization
- Smallest entry in the DeepSeek R1 distill series, making it the cheapest way to test R1-style prompting patterns before scaling up to larger distills
- Dense Qwen-based architecture is widely supported by open inference stacks such as vLLM, SGLang, llama.cpp and Ollama
- Small parameter count generally translates into high token throughput and low time-to-first-token on a given accelerator
- Useful as a draft model for speculative decoding alongside larger models in the same tokenizer family
- Available from multiple hosted providers as well as for self-hosting, so pricing and latency can be compared side by side in the table on this page
Limitations
- 1.5B parameters limits factual recall and general knowledge sharply compared with mid-size and large models
- Reasoning-style output is verbose, so a single answer can consume many more output tokens than a comparable non-reasoning small model
- Accuracy is well below the larger R1 distills (7B through 70B) and below DeepSeek-R1 itself on hard reasoning tasks
- Our benchmark record currently reports no measured throughput or time-to-first-token figures, so speed comparisons must come from provider-specific testing
- We do not track a confirmed context window or modality support for this entry; effective limits depend on the serving provider
- Prone to looping or unterminated thinking segments on prompts outside its distillation distribution, which can require output token caps
Key Features
•Reasoning-style generation with an explicit thinking segment preceding the final answer
•Approximately 1.5 billion parameters, the smallest tier of the DeepSeek R1 distill family
•Qwen-family dense transformer base fine-tuned on DeepSeek-R1 outputs
•Compatible with common open-source serving runtimes (vLLM, SGLang, llama.cpp, Ollama)
•Quantization-friendly size for single-GPU, laptop, or edge deployment
•Suitable as a speculative-decoding draft model for larger Qwen-tokenizer models
•Available both as a hosted API model and for self-hosted inference
About DeepSeek R1 Distill Qwen 1.5B
DeepSeek R1 Distill Qwen 1.5B is part of DeepSeek's R1 distillation series, in which reasoning traces generated by the larger DeepSeek-R1 model were used to fine-tune smaller dense base models from the Qwen family. At roughly 1.5 billion parameters, it is the smallest member of that distill lineup, sitting well below the 7B, 8B, 14B, 32B and 70B distills in capacity while inheriting the same chain-of-thought output style.
The model behaves as a reasoning chat model: it tends to emit an explicit step-by-step thinking segment before a final answer, which makes it verbose relative to conventional instruction-tuned models of similar size. Its strengths are concentrated in short mathematical and logical problems of the kind emphasized during distillation, rather than in broad world knowledge, long-document work, or nuanced writing — a 1.5B parameter budget places hard limits on factual recall. We do not currently track a confirmed context window, modality list, or tool-calling support for this entry, and serving limits vary between providers, so check the provider's own model card for the exact window offered.
In practice this model is used where a reasoning-style model is wanted at very low compute cost: local and on-device experimentation, batch pipelines over large volumes of short prompts, draft generation for speculative decoding, and research on reasoning distillation. Compared with its larger R1 distill siblings, it trades substantial accuracy for lower memory footprint and faster generation, and it is a common baseline in papers and blog posts studying how far reasoning behavior can be compressed.
Common Use Cases
This model suits workloads where a reasoning-style output format matters more than depth of knowledge, and where cost or hardware constraints rule out larger models. Typical fits include high-volume batch processing of short math or logic prompts, structured extraction and classification where a brief chain of thought improves consistency, on-device or offline assistants, classroom and tutorial demonstrations of reasoning-model behavior, and research into distillation and inference-time scaling. It also serves as a draft model in speculative decoding setups. For production tasks involving long documents, broad factual coverage, multi-step agentic tool use, or difficult competition-level problems, one of the larger R1 distills or DeepSeek-R1 itself is a more appropriate starting point.
Frequently Asked Questions
How much does DeepSeek R1 Distill Qwen 1.5B cost to run?
Costs depend on the provider you choose and on how the workload is billed — per-token serverless inference, dedicated endpoints, or GPU rental for self-hosting all price differently, and rates change over time. Use the pricing table on this page to compare current options. One thing to budget for regardless of provider: because the model emits chain-of-thought text, output token counts per request are typically higher than for a non-reasoning model of similar size.
What is DeepSeek R1 Distill Qwen 1.5B best used for?
Short-form reasoning tasks at high volume and low cost — math word problems, simple logic and step-by-step extraction — plus local and offline experimentation, and use as a draft model for speculative decoding. It is also a common baseline for research on reasoning distillation.
How does it differ from DeepSeek-R1?
DeepSeek-R1 is the large model that generated the reasoning data; this entry is a much smaller dense model fine-tuned on that data. It reproduces R1's chain-of-thought output style at a fraction of the parameter count, but with substantially lower accuracy and far less factual coverage.
How does it compare with the larger R1 distills, such as the 7B, 14B or 32B versions?
It is the smallest in the series, so it needs the least memory and generally generates fastest, while the larger distills are markedly stronger on hard reasoning and knowledge-dependent prompts. If quality is falling short on your task, moving up one tier in the same distill family is the usual next step.
Can it be self-hosted?
Yes. The R1 distills are dense models released as open weights and are supported by mainstream open inference runtimes; at roughly 1.5B parameters the model runs on a single modest GPU and, when quantized, on CPU-only machines. Compare hosted per-token pricing against GPU rental rates in the table on this page to see which is cheaper for your volume.
Does it support images or tool calling?
We do not track confirmed modality or function-calling support for this entry. Check the model card of the provider you plan to use, since serving configurations and supported API features differ between hosts.