What happened in AI today

1. MiniMax releases M3.1 Flash, a lower-cost model for agent tasks

   MiniMax · Models

   MiniMax has released M3.1 Flash, a Flash-tier model that it says is strong at agentic tasks, coding, computer use and cybersecurity. The company says the model's usage cost for agentic tasks is 70% lower than its predecessor, and that it runs nearly three times faster in online inference. It is priced at $0.15 per million input tokens and $0.60 per million output tokens, and it is available through the MiniMax API, M Plan and MiniMax Code. MiniMax says the model weights will be released later.

   • MiniMax M3.1 Flash: Redefining Agent Performance and Cost

1. DeepSeek upgrades V4.1 Flash to DeepSeek-V4.1-Flash-1017

   DeepSeek · Models

   DeepSeek says its V4.1 Flash model has been upgraded to DeepSeek-V4.1-Flash-1017, and its official web and app models now run the new version. The API call method and pricing are unchanged, and developers use the deepseek-flash name to reach the latest model. DeepSeek reports gains across benchmarks, including 42.3 on Terminal-Bench 4.0 and 77.0 on DeepSWE v1.1, and says the model also cut task runtime and output length.

   • DeepSeek-V4.1-Flash-1017 Update

More stories

1. Microsoft AI finishes pre-training MAI-Base-2, an 80B-active, 2.8T-parameter model

   Microsoft · Models

   Microsoft AI says it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters. The company says it outperforms every open-weight model it tested on held-out evaluations and is 4-12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, the strongest open base model. Microsoft says it trained the model on 16k GB200 GPUs over 38 days in August and September 2026, with 21% average MFU and 89% goodput. The model is now in post-training, and Microsoft says it plans to license the weights to trusted partners through Foundry, citing organic training data that does not involve distillation from other models.

   • Frontier Pre-training at Microsoft AI

2. Mistral releases Timestral, a family of time-series forecasting models

   Mistral AI · Models

   Mistral has released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. Mistral says the largest model ranks first among pre-trained, zero-shot models on the GIFT-Eval, TIME and fev-bench benchmarks while using about two-thirds of the parameters of the previous best. The models produce probabilistic, multivariate forecasts, accept past and known-future covariates, and run on a standard server CPU without a GPU. Mistral lists the license as modified MIT, has published a GitHub repository and a Hugging Face collection, and offers fine-tuning on private data through Mistral Forge. A technical report is still to come.

   • Introducing Timestral

3. IBM releases Granite Switch 4.1 with decision and speech adapters in one model

   Hugging Face · Models

   IBM's Granite Switch 4.1 models, 8B and 30B, combine chat, decision-making and speech-to-text in one model served from a single endpoint. They include a new Decision aLoRA for general decisions, a Guardian aLoRA for safety decisions, and Granite-Speech-5.0-Turbo-CTC, which IBM calls the fastest ASR model on the OpenASR Leaderboard. IBM says the combined model uses about the same GPU VRAM as the base model alone, and that adapters can reuse the base model's KV cache. IBM also says the Guardian adapter outperforms several named models on average on safety-related decision tasks.

   • Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

4. Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open weights image model

   Black Forest Labs · Models

   Black Forest Labs released FLUX 3 Image Klein, a 5.6B parameter open weights model for text-to-image generation, image editing and multi-reference generation with up to 10 reference images. The company says it ranks first among the open weights models in its internal comparison, ahead of larger models such as Qwen-Image 2.1 and Krea 2, and gains 91 Elo in text-to-image over FLUX.2 [klein] 9B. The INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, and the weights are on Hugging Face and in the BFL API, with API pricing starting at $0.009 per 1K image. The FLUX Kommunity License allows free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. Larger organizations need self-hosted licensing or the API.

   • FLUX 3 Image Klein

More stories

1. Liquid AI releases LFM2.5-24B-A2B, an open-weight reasoning model for on-device agents

   Hugging Face · Models

   LFM2.5-24B-A2B is a 24B-parameter reasoning model that uses only 2B active parameters per token. It thinks before answering, calls tools, and handles multi-step tasks in popular agent harnesses. The post says it scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, a model with about 50% more total and active parameters. It decodes 184 tokens per second on an Apple M5 Max and reaches over 21K output tokens per second on a single NVIDIA B300 GPU. The model and a base version are open-weight on Hugging Face, with support in llama.cpp, MLX, vLLM, SGLang, and ONNX.

   • Run fast, private agents on-device with LFM2.5-24B-A2B

1. Qwen releases Qwen-Image-2.1-Turbo, an 8-step open-weight image model

   Alibaba · Models

   Qwen has released Qwen-Image-2.1-Turbo, an accelerated version of Qwen-Image-2.1 for text-to-image generation and image editing. It generates in 8 denoising steps and uses the same 7B visual generation architecture. Qwen calls it the most balanced and cost-effective image model in the Qwen-Image series. The checkpoint loads in Diffusers with QwenImage21Pipeline and includes its recommended sampling schedule, so users do not need to set up the scheduler by hand. The weights are freely downloadable under Qwen's Research License, which allows research and evaluation only; commercial use requires a separate license from Qwen.

   • qwen-image-2.1-turbo
   • Qwen/Qwen-Image-2.1-Turbo

More stories

1. Microsoft releases a 3-bit MAI-Code-1.1-Flash that runs locally with full 256K context

   Microsoft · Models

   Microsoft AI says MAI-Code-1.1-Flash is now available to download and run locally in a version optimized with 3-bit quantization. The company says it keeps the full 256K context window and has coding performance comparable to the full-precision model on SWE-Bench Verified and Terminal-Bench 2.1. Local model calls carry no inference charges, and Microsoft recommends devices with more than 120GB of RAM. Experimental access in the GitHub Copilot app, Copilot CLI and Visual Studio Code is due by the end of the month. Microsoft first released MAI-Code-1.1-Flash on August 11, when GitHub began rolling it out in Copilot. At the time, Microsoft said it cost a quarter as much as the model it launched at Build in June, used 25% fewer tokens per task and scored 22% higher on Terminal-Bench 2.1 in Copilot CLI.

   • MAI-Code-1.1-Flash: Better, faster, at a quarter of the cost

2. TII releases Falcon-ASR, a speech model for Arabic and the Emirati dialect

   Hugging Face · Models

   The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-ASR, a 1.6 billion parameter speech recognition model for Arabic, with a focus on the Emirati dialect. TII says it averaged a 20.92% word error rate across the six test sets of the Open Universal Arabic ASR Leaderboard, against 23.17% for the best published result in a snapshot checked on 30 September 2026. On TII's internal Emirati evaluation, it scored 22.73% word error rate, the lowest among the systems TII compared. The model also handles English, French, Spanish and Portuguese with the same weights, and TII reports a 5.74% mean word error rate on seven English test sets. A demo is on Hugging Face, and TII says API access and native applications are planned.

   • Introducing Falcon ASR

More stories

1. Google releases EmbeddingGemma 2, an open multimodal embedding model

   Google · Models

   Google has launched EmbeddingGemma 2, a 740 million parameter model built on the Gemma 4 architecture. It maps text, code, images, audio and video into one shared embedding space and is released under the Apache 2.0 license, which allows commercial use. Google says it leads sub-1B multimodal embedders on benchmarks such as MTEB Code and MAEB, and that its MTEB Code score rose from 68.76 to 78.68 over the first EmbeddingGemma. Google says text-only use needs as little as 270M parameters, and with quantization the full model needs about 567MB of active RAM on a Pixel 11 Pro. Weights are on Hugging Face and Kaggle, with Model Garden availability coming soon.

   • EmbeddingGemma 2: an open, lightweight multimodal embedding model

2. TII releases Falcon-Emirati-7B, a model tuned for Emirati Arabic

   Hugging Face · Models

   The Technology Innovation Institute's Falcon team has released Falcon-Emirati-7B, a 7B-parameter model built on its Falcon-H1-Arabic family to understand and write Emirati dialect. The team says it trained on Emirati web text, Modern Standard Arabic material about Emirati culture, and synthetic dialect data built with glossaries and style rules. TII says the model scores 84.83% on Alyah, a 1,173-question Emirati-dialect benchmark, ahead of every Arabic and multilingual model it compared, including larger ones. In a separate open-ended test judged by an LLM, TII says the model scored 0.52 on dialect fidelity, against 0.05 or lower for ALLaM-7B-Instruct-preview, gemma-3-27b-it, Jais-2-8B-Chat and Fanar-2-27B-Instruct. The model is available on TII's chat platform, and the article names no license or download link.

   • Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance

1. Microsoft launches streaming transcription model and two voice models

   Microsoft · Models

   Microsoft AI released MAI-Transcribe-2-Streaming, which it says ranks no. 1 for accuracy on Artificial Analysis for both final and partial transcripts. The model returns real-time text in 60 languages and produces first partial results in just over 100ms, Microsoft says. It costs $0.54 per hour of audio through the end of the year. Microsoft also announced MAI-Voice-2.1, a text-to-speech model for 23 languages at $22 per 1M characters, and MAI-Voice-2.1-Flash at $15 per 1M characters with 150ms end-to-end latency. Both voice models support voice cloning from a few seconds of audio and have built-in consent guardrails, Microsoft says.

   • Our first streaming transcription model debuts at no. 1 on Artificial Analysis

More stories

1. Cohere launches Embed 5 embedding models in Pro and Fast tiers

   Cohere · Models

   Cohere released Embed 5, a new family of embedding models for enterprise search, RAG, and AI agents. Cohere says Embed 5 Pro delivers its strongest retrieval to date across visually rich documents, financial data, and parsed PDFs, while Embed 5 Fast offers competitive quality for latency- and cost-sensitive work. Both support text and image inputs, 100+ languages, and 128K-token context, share one embedding space so teams can index with Pro and query with Fast, and are available on the Cohere API, Model Vault, Microsoft Foundry, and Amazon SageMaker at $0.12 and $0.08 per million tokens.

   • Introducing Embed 5—A New Family of Frontier Embedding Models

1. OpenAI launches GPT-6 Sol and Luna as cheaper everyday models

   OpenAI · Models

   OpenAI says GPT-6 Sol and Luna bring GPT-6 Astra methods to faster lower cost models. Prices fall 50 percent to 2 dollars input for Sol and 0.10 dollars for Luna. They are available in ChatGPT Work and Codex and in the API.

   • Introducing GPT-6 Sol and Luna

2. Anthropic releases Claude Opus 5.5 with lower cost and stronger safety

   Anthropic · Models

   Anthropic says Opus 5.5 matches Claude Fable 5.1 on most work and costs 40 percent less to run than Opus 5. Pricing is 4 dollars input and 20 dollars output per million tokens, with faster output and higher usage limits.

   • Introducing Claude Opus 5.5

3. Google launches Gemini 3.8 voice models with custom voice design

   Google · Models

   Google introduced Flash TTS for creative voice design and Flash-Lite TTS for dubbing and agents. Google says users can create voices from prompts in over 100 languages and direct delivery line by line, with watermarking on output.

   • Gemini 3.8 text-to-speech says hello

4. Xiaomi releases MiMo V2.6 flagship, flash and speed models

   Xiaomi · Models

   Xiaomi says the V2.6 series includes a trillion-parameter omni-modal pro model, a low-cost flash model, and an ultraspeed version up to 20 times faster. The models target complex projects and real-time work.

   • Model Release

5. ElevenLabs releases medical transcription model for general use

   ElevenLabs · Models

   ElevenLabs says Scribe v2 Medical cuts errors on clinical audio by about 35 percent versus base Scribe v2. It is available via API and is HIPAA-eligible with retention controls.

   • Scribe v2 Medical is now available to everyone