Executive Summary
- Mistral has unveiled a preview of Mistral Large 4 ('le Chonk'), a 1-trillion-parameter natively multimodal open-weights model, with the full weight release scheduled for later this month.
- Perplexity debuted 'pplx-decider-v1.1-27b', an open-weights multimodal decision model topping Hugging Face's Decision Index 0.3 benchmark, alongside a 50% pricing reduction to $0.02 per million input tokens.
- Google launched EmbeddingGemma 2, an open, lightweight multimodal embedding model optimised specifically for on-device representation workloads.
- Supabase announced key Select 2026 platform upgrades, introducing automated backend generation from code, native Model Context Protocol (MCP) server support, and streaming Postgres data pipelines.
- OpenBMB quietly pushed MiniCPM-V 4.7 (35B-A3B) to Hugging Face, surfacing an undocumented vision-language model variant lacking formal licensing or usage parameters.
In Detail
Mistral AI
- Mistral Large 4 (ML4 / 'le Chonk') preview launched on Mistral Studio; 1T parameter multimodal model
- Open model weights scheduled for release at the end of October 2026
Hugging Face / AGI Hunt
- Perplexity releases pplx-decider-v1.1-27b open-weights model topping Hugging Face Decision Index 0.3
- Input pricing halved to $0.02 per million tokens
- Falcon-Emirati dialect LLM released by TII UAE
Google Developers
- EmbeddingGemma 2 released as an open, lightweight multimodal embedding model for on-device search and retrieval
Orca Router
- MiniCPM-V 4.7 (35B-A3B) released silently to Hugging Face by OpenBMB without model card or license
Supabase
- Select 2026 updates: Supabase from code, MCP server integration for agents
- Real-time Postgres streaming pipelines to BigQuery, Snowflake, and ClickHouse
- OrioleDB high-performance storage engine enters Public Beta
Model Releases
- Mistral Large 4 (ML4 / le Chonk) — Mistral AI. Mistral announced a preview API for Mistral Large 4, a 1-trillion-parameter natively multimodal model, with open weights slated for end-of-month release. Read more
- pplx-decider-v1.1-27b — Perplexity. Perplexity released an updated 27B open-weights multimodal decision model that ranks first on the Hugging Face Decision Index 0.3 benchmark at $0.02/1M input tokens. Read more
- EmbeddingGemma 2 — Google. Google introduced EmbeddingGemma 2, an open lightweight model designed for natively multimodal embeddings across text and visual inputs. Read more
- MiniCPM-V 4.7 (35B-A3B) — OpenBMB. OpenBMB uploaded an unannounced 35B parameter vision-language model variant to Hugging Face with an active MoE architecture. Read more
What to Watch
Interpretation, not reported fact. Each point follows from a source listed below — follow the links and judge for yourself.
- The arrival of 1-trillion parameter open-weight models like Mistral Large 4 will test whether enterprise private deployments can match proprietary hosted frontier systems in cost-efficiency. (Basis: Mistral Large 4 preview announcement and scheduled end-of-month open-weights drop.)
- Decision-specific routing models under 30B parameters are becoming viable low-latency dispatchers for agentic architectures as cost approaches near-zero levels. (Basis: Perplexity pplx-decider-v1.1-27b scoring top marks on HF Decision Index 0.3 at $0.02 per million input tokens.)
- Databases adopting standardized agent protocols (MCP) will accelerate autonomous database scaffolding and schema migrations handled directly by coding agents. (Basis: Supabase releasing native MCP server support and backend-from-code toolsets.)
This brief is generated automatically by an AI system from the sources listed above, and published without human editorial review. Summaries and any interpretation may contain errors — follow the source links before relying on anything here.
Sources
- Mistral AI (https://mistral.ai/news/)
- Hugging Face / AGI Hunt (https://agihunt.info)
- Google Developers (https://blog.google)
- Orca Router (https://www.orcarouter.ai)
- Supabase (https://supabase.com)