DeepSeek plans V4 multimodal model release this week, sources say
chips deepseek multimodal open-source reasoning startup
| Source: Mastodon | Original article
DeepSeek, the Chinese startup that released the open‑source R1 reasoning model earlier this year, is set to launch its first multimodal system, V4, within days. The new model will generate text, images and video from a single prompt, marking DeepSeek’s entry into a space dominated by models such as Qwen3.5‑Omni and Google’s Gemini‑3.1 Pro, which we covered in our March 31 benchmark roundup.
Sources close to the company say the rollout has been engineered together with Huawei and Cambricon to run efficiently on China’s home‑grown AI accelerators. By tailoring the architecture to the Ascend and MLU chip families, DeepSeek hopes to keep inference costs low while delivering competitive latency – a strategy echoed by Google’s recent TurboQuant memory‑saving claims. The partnership also signals a tightening of the Chinese AI supply chain, where software and silicon are increasingly co‑designed to reduce reliance on foreign hardware.
The announcement matters for several reasons. First, an open‑source multimodal model could democratise access to high‑fidelity video generation, a capability that has so far been confined to proprietary services. Second, DeepSeek’s chip‑level optimisation may set a new performance‑price benchmark for domestic AI workloads, potentially reshaping the economics of large‑scale deployment in China’s cloud market. Finally, the timing aligns with a wave of multimodal releases that are pushing the frontier of generative AI beyond static media, intensifying competition for research talent and ecosystem partnerships.
What to watch next: benchmark results from the 2026 Multimodal AI Benchmark will reveal how V4 stacks up against Qwen3.5‑Omni and Gemini‑3.1 Pro in accuracy, speed and cost. DeepSeek’s licensing terms and the availability of pre‑trained weights will indicate whether the model will stay truly open‑source or shift toward a commercial API. Finally, follow‑up statements from Huawei and Cambricon could hint at broader chip‑software bundles aimed at enterprises seeking in‑house generative AI capabilities.
Sources
Back to AIPULSEN