DeepSeek R1, released in January 2025, was DeepSeek’s first dedicated reasoning model — a 671-billion-parameter Mixture-of-Experts model (37 billion active per token) that matched OpenAI’s o1 on math, code, and reasoning benchmarks. It’s now over a year and a half old, and DeepSeek has moved well past it. Reasoning got folded into a hybrid architecture in August 2025, and the current flagship — DeepSeek V4 — had a major release just weeks ago.
Here’s what R1 actually was, what happened to it, and what’s current as of this writing.
What DeepSeek R1 Actually Was
R1 is built on the DeepSeek-V3-Base architecture: 671B total parameters, 37B activated per token (Mixture-of-Experts), with a 128K token context window. On release, it reported 79.8% on AIME 2024, 97.3% on MATH-500, a 2,029 Codeforces rating, 71.5% on GPQA Diamond, and 90.8% on MMLU — competitive with OpenAI’s o1 at a fraction of the cost. The official paper is at arXiv:2501.12948.
In May 2025, DeepSeek shipped R1-0528, a substantial reasoning upgrade — AIME 2024 jumped from 70.0% to 87.5%, and GPQA from 71.5% to 81.0%. That turned out to be R1’s last release under its own name.
What Happened to R1: The Timeline
DeepSeek didn’t retire R1 so much as absorb it. Starting in August 2025, reasoning stopped being a separate model line and became a mode within DeepSeek’s main models:
- August 2025 — DeepSeek-V3.1 introduced a hybrid architecture with both thinking and non-thinking modes in one model
- September 2025 — V3.1-Terminus (language-consistency fixes), then V3.2-Exp
- December 2025 — DeepSeek-V3.2 shipped, with a temporary V3.2-Speciale variant that hit Gemini 3.0 Pro-level results and gold-medal scores at the 2025 IMO and IOI
- April 2026 — The DeepSeek-V4 family launched (V4-Pro and V4-Flash), trained on over 32 trillion tokens with a native 1M-token context window
- July–August 2026 — V4-Flash went GA, followed by V4-Pro GA on August 13, then an experimental vision model, V4-Flash-Vision-Exp, on August 21
If you call the deepseek-reasoner API endpoint today, you’re not getting R1 — you’re getting the current model’s thinking mode, continuously upgraded in place since R1 first shipped.
DeepSeek V4 Pricing
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
| deepseek-v4-flash | $0.007–$0.014/1M | $0.22–$0.44/1M | $0.66–$1.32/1M |
| deepseek-v4-pro | $0.022–$0.044/1M | $0.66–$1.32/1M | $1.98–$3.96/1M |
| deepseek-v4-flash-vision-exp | $0.007–$0.014/1M | $0.22–$0.44/1M | $0.66–$1.32/1M |
Prices are per million tokens; the lower end of each range applies during off-peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday–Friday), which run at half the peak rate. All three models support a 1M-token context with up to 384K tokens of output.
Should You Still Use R1?
Not directly — there’s no reason to target R1 specifically anymore. If you’re calling DeepSeek’s API, you’re already on the current model whether you use deepseek-chat or deepseek-reasoner; both have been continuously upgraded since R1 launched. If you’re comparing DeepSeek to other providers for a coding or reasoning workload, evaluate V4-Pro (best quality, higher cost) or V4-Flash (cheaper, still strong) on your actual workload rather than benchmarks from an 18-month-old model.
Frequently Asked Questions
Q: Is DeepSeek R1 still available?
A: The R1 and R1-0528 checkpoints are still published on Hugging Face and GitHub, but DeepSeek’s hosted API no longer serves R1 directly — the deepseek-reasoner endpoint now points to the current model’s thinking mode.
Q: How many parameters does DeepSeek R1 have?
A: 671 billion total, with 37 billion active per token — it’s a Mixture-of-Experts model, not a dense 180B model as sometimes reported.
Q: What replaced DeepSeek R1?
A: Reasoning became a built-in mode of DeepSeek’s main model line starting with V3.1 in August 2025, culminating in the current DeepSeek V4 family (V4-Pro and V4-Flash), which had its most recent update on August 21, 2026.
Q: Is DeepSeek V4 open source like R1 was?
A: Check DeepSeek’s official Hugging Face page and licensing for the specific V4 checkpoint you want — release terms have varied across DeepSeek’s model line and shouldn’t be assumed to carry over from R1’s release.
Quick Summary:
- DeepSeek R1 (Jan 2025) was a 671B-parameter MoE reasoning model, not the 180B dense model earlier reporting claimed
- R1-0528 (May 2025) was its last release under the R1 name — a big reasoning jump (AIME 70.0% → 87.5%)
- Reasoning was folded into DeepSeek’s main model line starting with V3.1 (August 2025)
- The current flagship is DeepSeek V4 (V4-Pro and V4-Flash), with V4-Pro reaching general availability on August 13, 2026
- There’s no reason to target R1 specifically today — the deepseek-reasoner API endpoint has been continuously upgraded past it