12 stories about AI models, newest first. Each one links to its primary source.
Google DeepMind's EmbeddingGemma 2 maps text, code, images, video and audio into one 768-dimension embedding space. It has 740 million parameters in total, is released under Apache 2.0, and its encoders can be loaded selectively. The benchmark scores are Google's own.
On 7 October 2026 Google launched Playground, an experimental browser-based platform where US users aged 18 and over can create, play and share custom games by typing prompts. Creation access depends on the user's Google AI subscription tier. Google does not name the model behind it.
On 7 October 2026 Liquid AI released two open-weight decision models on Hugging Face: d1-3B (text and image) and d1-omni-600M, an experimental checkpoint (text with image or audio). Liquid says they produce an answer in a single forward pass, with d1-3B answering in under 50 ms on every measured device. Benchmarks are Liquid's own.
On 7 October 2026 Perplexity published two ColBERT-style late-interaction embedding models, pplx-embed-v2-late-9B and 0.6B, for text and image retrieval with a shared embedding space. Both are on Hugging Face; the post names no license or price. Benchmark numbers are Perplexity's own.
On 7 October 2026 the UAE's Technology Innovation Institute published Falcon-ASR, a 1.6 billion-parameter speech recognition model for Arabic with a focus on Emirati, plus English, French, Spanish and Portuguese in the same weights. TII reports a 20.92% average word error rate on six Arabic test sets; the post names no license.
On 7 October 2026 Microsoft AI released MAI-Code-1.1-Flash, which it says costs a quarter of MAI-Code-1.0 and uses 25% fewer tokens per task. It is live in GitHub Copilot and can be downloaded to run locally, with more than 120 GB of RAM recommended. Microsoft gives no dollar price, license or download location, and all figures are its own.
Xiaomi announced MiMo-V2.6 on 22 September 2026: a 1.02-trillion-parameter Pro model and a 309-billion-parameter Flash model, both multimodal, with weights on Hugging Face tagged MIT. Artificial Analysis lists Pro at 46 on its Intelligence Index; most other scores are Xiaomi's own.
Anthropic released Claude Haiku 5.5 on 7 October 2026. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, the same list price as OpenAI's GPT-6 Luna. The benchmark scores are Anthropic's own.
Reflection announced Beam on 5 October 2026, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters. The company says it will release the weights under Apache 2.0 later in October. The benchmark scores are Reflection's own, and in its table Beam trails several newer Chinese models.
Google released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash, on 6 October 2026. Its API price is $0.0336 per 1K image, about half the $0.067 of Nano Banana 2 at 1K and 2K, and about a quarter lower at 4K. The quality scores come from Google's own model card.
Mistral AI released Mistral Large 4 in public preview on 6 October 2026: a natively multimodal mixture-of-experts model with 1 trillion parameters, 49 billion of them active, priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the weights will follow by the end of October.
Between 28 September and 5 October 2026, OpenAI launched Dots agents and GPT-6.1 Sol but scrapped GPT-6.1 Astra after safety tests. Google released Gemini 4 Argon to cybersecurity partners only, six AI companies signed a voluntary White House accord, and Anthropic's IPO filing was reported.