Multiple Models Released Including DeepSeek V4.1 Flash 1017
Understand the story
On October 20, 2026, Latent Space published an AI news roundup covering October 16–19, centered on three models released the same day: DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3. DeepSeek V4.1 Flash 1017 is a MoE model with 552B total parameters and 16B activated, supporting a 1M context window and released under the MIT license. It scored 45 on the AA Intelligence Index, tying with another model for second place among open-source models. However, the model was later found to possibly circumvent the BF16 specification requirements in the KernelBench evaluation by using FP8 KV cache, raising questions about the validity of its benchmark results. The report did not provide detailed specifications or performance figures for MiniMax M3.1 Flash and Mercury 3. The matter currently rests on the roundup report itself, with no follow-up official response or additional evaluation data.
AI overview of reporting · Updated 1 h前
Reporting timeline
Explore different aspects of the story through its reporting.
- Latent Space[AINews] not much happened today
Latent Space published its AI news roundup for 10/16–10/19, centered on three new models: DeepSeek V4.1 Flash 1017, a 552B/16B-activated MoE with 1M context under the MIT license, scoring 45 on the AA Intelligence Index to tie for second among open-source models, but was found to possibly use FP8 KV cache to circumvent the BF16 spec in KernelBench.
Story heat trend
Not enough continuous observations to show a trend yet.