AI Daily · Oct 26, 2026: MiniMax M3.1 Flash free access ends early as 50% discount launches
MiniMax M3.1 Flash's free window closed early due to provider support changes, with two more weeks on 0G; B.AI runs a limited-time 50% discount on the same model. Kimi reset 25% of monthly quota for K3.1 Pro subscribers.
Pick a date
-
MiniMax M3.1 Flash free access ends early, stays on 0G
MiniMax said provider-side support for the free window changed, so free access to M3.1 Flash ends at 16:00 UTC rather than Oct 24; afterward it remains available on 0G Private Computer for two weeks. Unsloth also released 2-bit GGUF weights that keep 91.5% top-1 accuracy at 45% smaller size and run on 150GB RAM.
-
MiniMax-M3.1-Flash offered at 50% off across web, app and API
B.AI runs a limited-time 50% discount on MiniMax-M3.1-Flash (Space Bunny), a model aimed at agent tasks, software engineering and visual reasoning, from Oct 26 17:00 to Nov 2 16:59 SGT on web, Android app and API.
-
Moonshot Kimi resets 25% of monthly quota for K3.1 Pro subscribers
Moonshot AI apologized for rate limits and excess quota consumption on the K3.1 Pro subscription plan and gave every subscriber on every plan a one-time usage reset restoring 25% of their monthly quota, about one week of usage; usage below 25% resets to zero.
-
Qoder Lite coding agent plan costs $10 monthly for 1,000 credits
Qoder Lite is priced at $10 per month with 1,000 credits, doubled in the first subscription month, and includes the full agent harness, one credit pool across Desktop, IDE, JetBrains and CLI, plus new Qwen, DeepSeek, Zhipu, Kimi and MiniMax models.
-
Tencent open-sources Hy-SimulMT streaming speech translation model
Tencent released Hy-SimulMT, a self-hosted simultaneous Chinese-English speech translation model, under Apache 2.0 with six checkpoints including FP8 and GGUF. It keeps latency around 3 seconds on the RealSI long-form benchmark and scores 84.7 on RealSI-Sentence.
-
MiniMax H3 video model gains 8-step reference control and 4-step distillation
Community contributions to the open-weights MiniMax H3 video generation model add 8-step reference-conditioned video with audio, 4-step distillation, sparse attention and reference-token compression, with serving on GPU servers, TPUs and Apple Silicon.
-
2-step distilled LoRA for Qwen Image 2.1 runs realtime on ZeroGPU
A community release ships a 2-step distilled LoRA (DMD2, rank 32) for Qwen Image 2.1 Realtime that runs in realtime on ZeroGPU, with quality described as competitive with 8-step Turbo and 40-step Base.
-
Qwen open-sources Babel-themed research on agent communication breakdowns
Qwen Developers released open-source research named after the Tower of Babel, examining where agent understanding breaks down, including tool arguments and database entities, with a paper, GitHub code and datasets.