Rumors Swirl of Imminent DeepSeek and GPT-5.3 Releases
Recent leaks and speculation suggest two major AI model releases may be imminent. A new video discusses rumors about the potential launch of DeepSeek V4 Lite and an iterative update from OpenAI, reportedly named GPT-5.3 or “Garlic.” While unconfirmed, the anticipation points to the rapid pace of development in the foundation model landscape.
- Leaks suggest GPT-5.3, codenamed "Garlic," was born from an internal "Code Red" at OpenAI to counter rapid advances from competitors like Google's Gemini 3 and Anthropic's Claude 4.5. - The "Garlic" codename for GPT-5.3 reportedly reflects a strategic shift from building bigger models to creating "smarter and denser" ones, focusing on computational efficiency to pack more reasoning into a faster, more cost-effective architecture. - Projections for GPT-5.3 indicate it could surpass the 83.7% human baseline on SimpleBench, a benchmark for common-sense reasoning, and leaked specifications include a 400,000-token context window. - DeepSeek's V4 models are expected to feature a novel "Engram" architecture that separates static memory from reasoning, allowing factual knowledge to be stored in more cost-effective CPU RAM instead of GPU VRAM, potentially lowering deployment costs for startups. - The rumored release for DeepSeek V4 is around mid-February 2026, coinciding with the Lunar New Year, a strategy the company has used previously to maximize visibility. - Leaked examples from DeepSeek V4 Lite showcase a specialized ability to generate efficient Scalable Vector Graphics (SVG) code, producing a detailed rendering of an Xbox controller in just 54 lines. - DeepSeek's prior model, DeepSeek-Coder-V2, already demonstrated performance comparable or superior to closed-source models like GPT-4 Turbo and Claude 3 Opus on coding and math benchmarks, indicating a strong trajectory for the V4 release.
