New AI Model Releases from Kimi K3 to Fish Audio

모델 출시 | Wed Jul 29 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 5 sources

New AI models were unveiled across various domains including open-weight LLMs, voice generation, Earth observation, and encoder models.

Analysis

[Kimi] unveiled Kimi K3, a 2.8T parameter open-weight model ^[2]

Currently the largest open-weight model
Scales Kimi Linear architecture from 48B to 2.8T
Introduces LatentMoE and multi-head latent attention
Removes RoPE from all layers and uses NoPE
Native multimodal support

[Deltafin] released an experimental project running Kimi K3 on M1 Max ^[3]

Runs a 2.8T MoE model on a 64GB laptop
Approximately 16 seconds per token on M1 Max
--full mode requires 1.7TB of disk
--stream mode can start with 215GB
Supports MXFP4 kernels and int8 spine optimization

[Fish Audio] raised $52M seed round to develop AI voice models ^[1]

Led by Coreline Ventures and Capital Today
Library of over 15
000 natural language controls
Over 8 million users acquired
Reached $21M in ARR
Released 5 models in the past year

[Ai2] launched OlmoEarth Platform for Earth observation foundation models ^[4]

Pretrained on approximately 10TB of multimodal satellite data
Used for deforestation monitoring
food security
and wildfire risk analysis
Processes continental-scale inference in about a day
Processes at a cost of less than a few pennies per km²
Supports everything from fine-tuning to large-scale inference

[Liquid AI] released LFM2.5-Encoders optimized for long-context CPU inference ^[5]

Two model sizes: 230M and 350M
Supports 8
192 token context
About 3.7x faster on CPU than ModernBERT-base
Excellent on GLUE
SuperGLUE
and multilingual tasks
Applicable to intent router
PII detection
and more

Sources