Mistral launches public preview of Mistral Large 4
Mistral has launched a public preview of Mistral Large 4, a 1 trillion-parameter natively multimodal model with 49 billion active parameters. The company claims the model significantly outperforms open-weight models developed in the US or Europe. Mistral plans to release the model weights by the end of the month. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters. The reinforcement learning run behind this preview is still in flight, with the model showing no signs of saturation.
mistral.ai
·
HN discussion
Polars 2.0 releases with out-of-core support and first-class SQL
The Polars team has released version 2.0, introducing out-of-core support and first-class SQL capabilities. The update enables spill-to-disk by default, starting at approximately 80% of available RAM, to handle high-memory workloads. SQL support has expanded significantly, with the project claiming Polars now leads DataFusion and DuckDB in TPC-H and TPC-DS1 benchmarks. Additionally, the release adds a new Map dtype and stricter data type enforcement. While the team notes a scaling issue at high core counts, they state they have diagnosed the cause and plan to fix it in the next release.
pola.rs
·
HN discussion
Gleam v1.19.0 switches from generating Erlang source to abstract forms
Gleam v1.19.0 changes how the language compiles for the Erlang virtual machine. Giacomo Cavalieri rewrote the code generator to output Erlang abstract forms instead of source code. These forms are a metadata-annotated tree representing Erlang syntax, normally produced by the Erlang parser. By encoding them in Erlang's external term format, the compiler loads generated code directly and skips the front half of the Erlang compiler pipeline.
gleam.run
·
HN discussion
OpenTPU: An AI-designed accelerator runs modern LLMs on a Kintex-7 FPGA
The openTPU project presents an open-source AI accelerator designed by AI agents, consolidating RTL, ISA, simulator, compiler, and profiler into a single monorepo. The design runs ten modern models with real weights on an Inspur YPCB-00338 card equipped with a Xilinx Kintex-7 xc7k480t FPGA. Measured performance reaches 14.5 GB/s, achieving 85% of peak DRAM bandwidth. Recent builds decode LFM2-2.6B, SmolLM3, and Phi-4-mini 8-9% faster than previous versions, operating at 91-94% of DRAM peak efficiency.
github.com
·
HN discussion
Global study finds ecosystems have far less spare capacity than assumed
A global analysis published in Nature Ecology & Evolution finds that ecosystems have far less spare capacity than previously thought. The study, led by King's College London and Imperial College London, combined 423 published studies to show that most ecosystem benefits rise steadily with biodiversity rather than leveling off. This challenges the assumption that functional redundancy allows species to substitute for one another. Ocean carbon sequestration showed the strongest response to diversity, though authors note this result rests on limited data. Conversely, protection against coastal hazards often depends on specific foundational species, suggesting conservation must balance overall diversity with safeguarding irreplaceable species.
phys.org
·
HN discussion
Fervo Energy begins selling power from first enhanced geothermal plant
Fervo Energy began selling electricity from its Cape Station plant on September 30, one day ahead of schedule. This marks the first time an enhanced geothermal company has reached a key commercial milestone. The initial block, which took 23 months to build, now supplies one-third of the site's eventual 100-megawatt capacity. Fervo aims to reduce construction time for future blocks to 18 months. Google and Southern California Edison have committed to purchasing power from the Utah-based facility.
techcrunch.com
·
HN discussion