Liquid AI releases LFM2.5-24B-A2B, a 2B-active open-weight MoE built for on-device agents
Liquid AI released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model that activates about 2 billion parameters per token. It is a mixture-of-experts design with a hybrid of short convolutions and grouped-query attention and a 131,072-token context window. Liquid reports it matches or beats Qwen3.6-35B-A3B on five of eight agentic and tool-use benchmarks and runs at 184 tokens/s on an Apple M5 Max; weights are open under Liquid's model license, with API pricing of $0.10 input / $1.00 output per million tokens. The benchmark figures are Liquid's own and not independently verified.
Sources & depth
TensorRT-LLM v1.4.0rc0 adds GLM-5.3-Flash and MiniMax-M3 support, with known SM120 crashes
NVIDIA's TensorRT-LLM 1.4.0 release candidate adds support for GLM-5.3-Flash, K-EXAONE-2, Qwen-Image-Layered and MiniMax-M3 with NVFP4 KV cache, plus faster small-batch MoE routing and tiled FP8 context attention on SM120 GPUs. It also removes the Python backend of KVCacheManagerV2, a breaking API change. Listed known issues include TP2 worker-initialization hangs on H100 and illegal-memory-access crashes for DeepSeek MLA models on SM120/SM121 with CUDA graphs. It is a pre-release.
Sources & depth
Ollama v0.40.5 ports its MLX Metal kernels to CUDA
The Ollama v0.40.5 pre-release ports MLX's custom Metal kernels to CUDA, bumps the MLX version, and lets the OpenAI-compatible Responses API accept custom tool calls in conversation history. The desktop app also gets clearer cloud payment errors. The CUDA port is the notable part: it moves Ollama's MLX path beyond Apple hardware.
Sources & depth
- github.comv0.40.52026-10-15 01:15 IST
vLLM ships v0.32.0rc1 and proto-v0.6.0
vLLM published the v0.32.0 release candidate, whose headline change updates the Humming dependency and removes an NVFP4 scale workaround, alongside a proto-v0.6.0 tag. The release notes are terse; this is incremental rather than a major release.
Sources & depth
Anthropic opens a Singapore office and signs an AI safety MOU with the government
Anthropic opened its fifth Asia-Pacific office in Singapore and signed an MOU with Singapore's Ministry of Digital Development and Information covering AI safety, security and model evaluation, adoption research and responsible-use work. Anthropic says Claude already powers services at GovTech Singapore, Grab and DBS.
Sources & depth