Releases·Model Release
Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax
A YouTube interview with Thomas Wolf and Olive Song of MiniMax explains the design of MiniMax M3. The model pairs a functional one-million-token context window with coding, agentic, and multimodal capabilities, addressing agents' need to work across long conversations, tool responses, and complex environments. An intern designed the sparse-attention architecture behind it.
18 engagement1 sourceFri, Sep 4, 2026, 01:00 PM
AI Overview
- An intern at MiniMax designed the sparse-attention architecture behind MiniMax M3's one-million-token context window, Olive Song says.
- MiniMax trained M3 as multimodal from its first training step, letting one model process text, images, and video.
- M3 already helps MiniMax build its successor M3.1, and the company's products serve 300 million users.
- MiniMax uses internal agent harnesses to automate parts of its AI research workflow.
Related
Social·Community
How MiniMax M3 Was Built: Sparse Attention and Native Multimodality — Olive Song
At AI Engineer World's Fair 2026, Olive Song, RL lead at MiniMax, presented the architecture behind MiniMax M3, an open-weight model with 1M-token context, native multimodality, and frontier-level coding and agentic ability. She detailed MSA, MiniMax's sparse attention design, and argued training jointly on text and vision from the first step outperforms bolting vision on later.
At AI Engineer World's Fair 2026, Olive Song, RL lead at MiniMax, presented the architecture behind MiniMax M3, an open-weight model with 1M-token context, native multimodality, and frontier-level coding and agentic ability. She detailed MSA, MiniMax's sparse attention design, and argued training jointly on text and vision from the first step outperforms bolting vision on later.
21 engagement1 sourceyoutubeSat, Oct 10, 2026, 10:30 PM
News·Company News
MiniMax RL lead Olive Song details open-weight agent stack
Olive Song, who leads reinforcement learning at MiniMax, discusses the company's open-weight models and the infrastructure behind them, emphasizing a belief in open source. She covers agentic coding, computer use trained with RL against environments like OS World, and the unglamorous engineering that makes launches real.
Olive Song, who leads reinforcement learning at MiniMax, discusses the company's open-weight models and the infrastructure behind them, emphasizing a belief in open source. She covers agentic coding, computer use trained with RL against environments like OS World, and the unglamorous engineering that makes launches real.
9 engagement1 sourceyoutubeFri, Jul 31, 2026, 03:00 AM
Releases·Model Release
Moonshot details Kimi K3 training in Inside Kimi K3 Part 3
Moonshot AI published Part 3 of its Inside Kimi K3 series, explaining how it trained the architecture described in earlier parts into a multimodal, million-token, reasoning and agentic model. The post focuses on the training methodology behind Kimi K3.
Moonshot AI published Part 3 of its Inside Kimi K3 series, explaining how it trained the architecture described in earlier parts into a multimodal, million-token, reasoning and agentic model. The post focuses on the training methodology behind Kimi K3.
0 engagement1 sourcerssWed, Aug 19, 2026, 07:34 PM
Social·Community
Users ask about MiniMax-M3 agentic workflow coherence
A developer asks the community about MiniMax-M3's practical behavior in long-running agent workflows, specifically whether it stays coherent across many tool calls or drifts as tasks grow. The question highlights interest in the model's agentic reasoning, tool use, coding, and large context window.
A developer asks the community about MiniMax-M3's practical behavior in long-running agent workflows, specifically whether it stays coherent across many tool calls or drifts as tasks grow. The question highlights interest in the model's agentic reasoning, tool use, coding, and large context window.
8 engagement2 sourcesredditWed, Aug 12, 2026, 10:14 PM
Releases·Tool Release
SGLang v0.5.17 adds day-0 support for Kimi K3, a 2.8T-parameter LatentMoE model
SGLang v0.5.17, released August 8, 2026, includes day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE model with 896 experts and 1M-token context. The release integrates advanced serving features like DCP, DSpark speculative decoding, and KDA-aware prefix caching, verified on NVIDIA GB300 and AMD MI35x.
SGLang v0.5.17, released August 8, 2026, includes day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE model with 896 experts and 1M-token context. The release integrates advanced serving features like DCP, DSpark speculative decoding, and KDA-aware prefix caching, verified on NVIDIA GB300 and AMD MI35x.
0 engagement1 sourcegithubSat, Aug 8, 2026, 12:19 AM