AI News

‹Tue, October 6›

01103

Mistral launches Mistral Large 4 preview, a 1T-parameter open-weight model

Mistral AI opened a public preview API for Mistral Large 4 (nicknamed "le Chonk"), which it calls the strongest open-weight model from the US or Europe. The natively multimodal model has 1 trillion total parameters and 49 billion active ones, was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, and its weights are due by the end of the month. On Artificial Analysis's Intelligence Index it scores 38, ahead of GLM-5.2 but well behind Claude Opus 5.5 at 58.

MMistral AI News·1d

0227

Models

Google releases Nano Banana 2.1 image model, halving output prices

Google released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash, adding mask-based local editing and stronger subject consistency — up to 14 reference images, keeping four characters and ten objects consistent — plus Google Search grounding and minimal/medium/high thinking levels. Pricing is roughly halved versus Nano Banana 2: a 1K image drops from 6.70 to 3.36 cents and a 4K image from 15.10 to 7.56 cents, while input and text/thinking output prices rise. It is available now in AI Studio (gemini-nano-banana-2.1), the Gemini app, Search AI Mode, Flow and other surfaces, and the Nano Banana 2 API retires on October 29. The Decoder notes 2.1 beats Nano Banana Pro on some benchmarks, but Pro often still produces better images in practice.

GGoogle AI Studio·1d

0334

Models

Google releases EmbeddingGemma 2, a 740M on-device multimodal embedding model

Google announced EmbeddingGemma 2 on October 6, a 740M-parameter open model built on Gemma 4 and released under Apache 2.0 that maps text, code, images, video and audio into one shared 768-dimensional space. It is modular: text/code alone needs 270M parameters, with optional vision (440M) and audio (570M) encoders up to 740M for full multimodal, and Matryoshka Representation Learning lets developers truncate vectors to 512, 256 or 128 dimensions for up to 6x storage savings. Google says it scores 78.68 versus 68.76 for the previous version on MTEB (Code) and beats some rival models twice its size; with quantization it uses about 191MB of RAM for text-only on a Pixel 11 Pro. Google also showed an experimental Mac app, AI Edge Foresight, for offline note-taking and personal knowledge retrieval.

GGoogle Developers Blog·2d

046

Models

Musubi releases open-weight PolicyLM-1.7B for content moderation

Musubi announced PolicyLM-1.7B, an open-weight decision model designed for real-time content moderation. The company says it can apply plain-English policies to messages in under 50 milliseconds without retraining when policies change, but these performance and flexibility claims come from the product announcement.

TTechCrunch AI·20h

055

Models

Claude Opus 5.5 composes retro game music in Scrimshaw Jukebox test

On his blog, Simon Willison tested whether Claude Opus 5.5 could compose music, asking it to first design a simple text-based music format, then build an artifact that plays it aloud with example tracks, aiming for the quality of the original Secret of Monkey Island. The result was "Scrimshaw Jukebox," a pixel-art browser player containing six original adventure-game tracks (Moonlit Harbor, The Ghost Galleon and others), written as plain text and performed by an in-browser synthesizer with a piano-roll score view and editing. Willison called the output "surprisingly good" while noting it leaned harder into the Monkey Island theme than he intended.

SSimon Willison's Weblog·1d

064

Models

Utopai X Reportedly Ranks Second in Artificial Analysis Video Test

Utopai Studios’ Utopai X reportedly scored 1,150 (±10) in an Artificial Analysis blind video evaluation, ranking second globally and first in the United States. The article says the model was post-trained from MiniMax H3 for film production and led the evaluation’s audio-sync and physics categories, but these claims are presented through a promotional report and are not independently verified here.

新智元·1d

That's everything.