Unverified50% confidenceFactExact time
DeepSeek-V4.1-Flash is based on multimodal pre-training with 45T tokens, supports context reasoning intensity continuously adjustable from 1 to 100, and uses the MIT license.
1
Sources
50%
Confidence
Long-term
Relevance
10/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDeepSeek and Huawei jointly optimized the Ascend 950 large-scale cluster solution, performing in-depth tuning for multi-card training and large model inference.74% similarVerifiedDeepSeek was incubated by quantitative private equity giant High-Flyer, and its R1 series models achieve reasoning capabilities close to GPT-4 level at extremely low training costs.73% similarUnverifiedDeepSeek等国产模型成本突破得益于MoE稀疏架构、FlashAttention注意力加速技术和高效分布式训练调度策略73% similarUnverifiedDeepSeek-V4.1-Flash is open-sourced under the MIT license, natively supports multimodal vision, and has a context length of 1 million tokens.73% similarUnverified微软的 DeepSpeed 库与 ORT Training 有深度集成，两者结合可在千卡规模集群上训练百亿参数级别模型71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/970667API
curl https://kongchang.com/api/v1/knowledge/claims/970667MCP
get_claim(id=970667)