Liquid AI LFM2.5: On-Device MoE With 1.5B Active Params
Liquid AI released LFM2.5-8B-A1B, an on-device MoE model with 8.3B parameters but only 1.5B active at inference. 253 tok/s on M5 Max, 128K ...
DeepSeek V4: MIT Open-Source Model Matches Claude and GPT-5.5
DeepSeek V4-Pro matches Claude Opus 4.7 on SWE-bench at $1.74/M tokens under an MIT license. ...
Apple Foundation Models: On-Device Swift AI, No API Key
Apple's Foundation Models framework lets Swift developers add on-device AI to iOS 26 apps with ...
vLLM v0.21.0: Spec Decode for Reasoning Models — Upgrade Now
vLLM v0.21.0 ships thinking-budget-aware speculative decoding, KV offload + HMA integration, and a Blackwell MLA ...
DeepSeek V4-Pro: Open-Source 1.6T Model — What Developers Must Know
DeepSeek V4-Pro scores 80.6% on SWE-bench at $3.48/1M tokens vs Claude Opus 4.6 at $25. ...
Sakana AI’s 7B Model Beats GPT-5 by Telling It What to Do
Sakana AI's RL Conductor, a 7B model trained with RL, outperforms GPT-5 on GPQA and AIME by orchestrating frontier models. Now available in ...
Qwen3-Coder-Next: Run a Frontier Coding Agent Locally
Qwen3-Coder-Next scores 70%+ on SWE-Bench and runs on a single RTX 5090 for ~$2,500. Here ...
Gemma 4 MTP: How Google’s 3x Inference Boost Works
Google released Multi-Token Prediction drafters for Gemma 4 on May 5, delivering up to 3x ...
9Router Tutorial: Eliminate AI Coding Rate Limits (2026)
Eliminate AI coding rate limits with 9Router. Route Cursor, Claude Code, Copilot through 60+ providers. ...
InsForge: Backend Platform for AI Coding Agents (Tutorial 2026)
InsForge is a Postgres-based backend platform built for AI coding agents. 1.6x faster than Supabase ...