DeepSeek V4
- Lab
- deepseek
- Weights
- open_weights
- Released
- 2026-08-24
- Context window
- 256,000
- Modalities
- text
- API
- yes
- Local-runnable
- no
- Status
- current
announcementmodel cardpricing
Scores
Coding
| Value | Source | As of | Reported by | Notes |
| 71.5 percent | swe-bench-verified ↗ | 2026-08-24 | labL | Seed placeholder. Lab-reported at release; independent run pending. |
Output speed
| Value | Source | As of | Reported by | Notes |
| 84 tokens_per_second | artificial-analysis-speed ↗ | 2026-08-24 | aggregatorA | Seed placeholder — first-party API. |
Cost
| Value | Source | As of | Reported by | Notes |
| 2.2 usd_per_million_blended | artificial-analysis-pricing ↗ | 2026-08-24 | aggregatorA | Seed placeholder. |
Runs well locally on 64GB Apple Silicon
| Value | Source | As of | Reported by | Notes |
| no | owner-notes ↗ | 2026-08-25 | independent | Open weights but ~600B MoE — does not fit 64GB even at 4-bit. |
Signals
- hf_downloads_30d: 184,000 as of 2026-08-25 source
- github_stars: 21,400 (+3800/7d) as of 2026-08-25 source
Verdicts
Events
3 DeepSeek's API documentation lists deepseek-v4-flash-vision-exp as a new experimental vision-capable model in the V4 line. Latent Space's AI News recap reports the corresponding weights are now open, with curator commentary noting the release brings DeepSeek to vision parity with Moonshot and GLM.
2026-08-29 Open weights Release DeepSeek API news / changelog 4 DeepSeek released V4 with full weights on Hugging Face under MIT. Lab-reported numbers put it within five Artificial Analysis index points of the closed frontier, which would be the smallest open/closed gap on record. Too large to run on consumer hardware, but hosted pricing undercuts frontier models by roughly 8x.
2026-08-24 Release Open weights Lab strategy DeepSeek V4 release notes 3 DeepSeek updated its hosted checkpoints (deepseek-v4-flash to DeepSeek-V4-Flash-0731, deepseek-v4-pro to DeepSeek-V4-Pro-0813, both accessible under the same model names), released an experimental image-input variant (deepseek-v4-flash-vision-exp), and put a new 'DeepSeek Harness' agent harness into developer preview for agent-tool builders.
← Back to Models