ERNIE 5.5 Flash Base is now open source, a 280B-parameter model with 14B active per token, and it already matches SOTA at its size. The full ERNIE 5.5 Flash is on its way.
Key points
Open source release: ERNIE 5.5 Flash Base is now open source.
Model size: 280B parameters with 14B active per token.
Sources
JITEngine custom Apple Silicon engine up to 2.7× faster
JITEngine offers custom inference engines built with SkyDiscover for Apple Silicon, claiming up to 2.7× faster performance than the fastest alternatives tested. It is presented as a just-in-time synthesis solution for LLM inference.
Key points
Engine: custom engines built with SkyDiscover for Apple Silicon
Performance: up to 2.7× faster than the fastest alternatives we tested
Sources
- Original post
- jitengine.ai - Linked in the post
frax.ik high-performance humanoid retargeting library announced, >100k FPS
frax.ik is a new CPU-only library for humanoid retargeting and inverse kinematics that runs in low microseconds. It claims orders-of-magnitude speedup over existing alternatives and avoids foot skidding, floor penetration, and self-collisions. The accompanying frax release reports IK pipelines achieving over 100,000 frames per second.
Key points
Runs entirely on CPU with compute times in the low microseconds range (~25-100 kHz).
IK and retargeting pipelines deliver performance above 100,000 FPS.
Sources
- Original post
- frax: Fast Robot Kinematics and Dynamics in JAX - Linked page
GPT-Live outbound telephony API GA launch
GPT-Live outbound telephony API is now GA, letting developers place calls from Live voice agents via any SIP trunk provider. Calls support HD Voice and can send and receive DTMF keypad events.
Key points
Release: GA for outbound telephony API
Features: HD Voice and DTMF support, works with any SIP trunk provider (e.g., Telnyx, Twilio)
Sources
dots iOS update adds inline reply, share, action
dots adds inline reply from iOS notifications, share to dot, and open dot from iPhone Action in the latest app release. The team says they are laser-focused on making dots faster and more reliable.
Key points
Inline reply: you can now reply to your dot inline from a notification on iOS
Share: share anything directly to your dot in iOS
Action: open your dot from the iPhone Action
Sources
Ripple-65B Preview release new 65B dense LLM
Ripple-65B Preview is a 65-billion-parameter dense LLM released today. It claims state-of-the-art performance on U.S. open-weight agentic coding tasks, surpassing larger models such as Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron.
Key points
Release: Ripple-65B Preview, a 65B dense LLM pretrained from scratch.
Claim: Sets new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing larger models listed.
Sources
Llama adds d1-Omni multimodal and d3-flash support
Llama now supports d1-Omni by liquidai for multimodal queries and d3-flash by vLLM. Update Llama today.
Key points
d1-Omni by liquidai enabled for multimodal queries
d3-flash by vLLM added, leading Jev Decision Index
Sources
PyTorch MoE grouped execution reduces kernel launches
PyTorch operations can be performance optimized for MoE models by grouping expert execution to cut kernel launches. The Nvidia blog uses biological foundation models as an example of efficient MoE training.
Key points
Replacement: MoE models replace a single dense feed-forward block with multiple expert networks.
Optimization: Grouped execution submits all expert work together, reducing launch and scheduling overhead.
Sources
- Original post
- bit.ly - Linked in the post
