Will DeepSeek release a model named R2 or V5 before 31 March 2027?
deadline 2027-03-31
confidence medium
tier 4 evidence
68%Asturna estimate
Resolution criterion
Yes if DeepSeek publishes on Hugging Face or api-docs.deepseek.com a model whose official name contains "R2" or "V5". Variants of V4 (V4 Pro, V4 Flash, V4-Thinking) do not count.
Verified starting state
DeepSeek V4, V4 Pro and V4 Flash were released on 24 April 2026. No R2 has ever been released despite rumours since 2025. DeepSeek has a history of unpredictable version jumps.
What pushes it up
- The reasoning model line (R1) has not been updated since January 2025; a successor (R2) is highly likely within the 27-month window ending March 2027 [Starting State]
- The 11-month window between the V4 release (April 2026) and the deadline is consistent with the 4-7 month gaps seen in earlier versions (V1 to V3) [Starting State]
- DeepSeek is reportedly raising 80 billion RMB to fund R&D and IPO preparations, which would provide massive capital for training next-generation models [0, 1, 2] (Note: This Tier 4 report is not yet corroborated by Tier 1 or 2 sources)
- Heavy recruitment for the elastic computing team suggests a significant scale-up in training infrastructure [4]
What pushes it down
- The development cycle between V3 and V4 was 16 months, suggesting that a V5 release within 11 months of V4 may be too aggressive if the trend of increasing complexity continues [Starting State]
- Resolution criteria specifically exclude V4 variants (Pro, Flash, Thinking), which are the most likely intermediate releases before a full V5 [Resolution Criterion]
What would move this number most
Whether DeepSeek prioritizes the 'R2' reasoning model or the 'V5' general model, as the 11-month window from the previous major general release is relatively tight.
How this number is built
| Deadline | 2027-03-31 |
| Base rate | 65% |
| Best evidence | tier 4 · 4 article(s) used |
| Independent runs | 65 · 70 (spread 5 pts) |
| Confidence | medium |
| Evidence cap | not triggered |
| Last revised | 2026-10-08 09:06 |
| Model | gemini-3-flash-preview |
Sources consulted
Probability history
2026-10-08 09:06 68% 10 article(s) on file · gemini-3-flash-preview
- What pushes it up. The reasoning model line (R1) has not been updated since January 2025; a successor (R2) is highly likely within the 27-month window ending March 2027 [Starting State]
- What pushes it up. The 11-month window between the V4 release (April 2026) and the deadline is consistent with the 4-7 month gaps seen in earlier versions (V1 to V3) [Starting State]
- What pushes it up. DeepSeek is reportedly raising 80 billion RMB to fund R&D and IPO preparations, which would provide massive capital for training next-generation models [0, 1, 2] (Note: This Tier 4 report is not yet corroborated by Tier 1 or 2 sources)
- What pushes it up. Heavy recruitment for the elastic computing team suggests a significant scale-up in training infrastructure [4]
- What pushes it down. The development cycle between V3 and V4 was 16 months, suggesting that a V5 release within 11 months of V4 may be too aggressive if the trend of increasing complexity continues [Starting State]
- What pushes it down. Resolution criteria specifically exclude V4 variants (Pro, Flash, Thinking), which are the most likely intermediate releases before a full V5 [Resolution Criterion]
- What would move this number most. Whether DeepSeek prioritizes the 'R2' reasoning model or the 'V5' general model, as the 11-month window from the previous major general release is relatively tight.
2026-09-24 09:08 39% 0 article(s) on file · openai/gpt-oss-20b (Groq)
Reasons not recorded for this revision (detailed history has only been kept since recently).