AI Briefing#reasoning
The latest AI and developer news about #reasoning, with the original source and a short summary.
The latest AI and developer news about #reasoning, with the original source and a short summary.
Zyphra has released ZAYA1-8B, a 760M active parameter MoE model.
It points out that claims about LLM limitations—stripped of the assumptions behind the mathematical theorems—have been exaggerated and consumed as such.
llm-echo 0.5a0, which provides a fake echo model for LLM, has been released.
PORTool improved both accuracy and tool-call efficiency together through step-wise importance estimation.
Manifest Router has begun supporting routing for GPT-5.5 and GPT-5.5 Pro.
A synthetic fine-tuning dataset of 8,706 examples based on Claude Opus 4.6/4.7 has been released.
GLM-5V-Turbo is an agent model that integrates multimodal perception, tool use, and end-to-end verification.
Mistral announced Medium 3.5, a 128B open weights multimodal model, in public preview.
SenseNova-U1 has been unveiled as a native multimodal model that unifies understanding, reasoning, and generation.
NVIDIA has unveiled the multimodal Nemotron Nano Omni 30B A3B Reasoning.
In an M3 Pro benchmark, Nemotron 3 Nano took first place among 4B-class models.
A 23-year-old amateur solved a 60-year-old Erdős problem using ChatGPT Pro's GPT-5.4 Pro.
GPT-5.5 topped the intelligence index on Artificial Analysis, but it also ranked among the most expensive.
Only Qwen, which has a long pre-saturation phase, succeeds at RLVR generalization under weak supervision, while Llama fails.
Shared a setup for running Qwen3.6-35B-A3B with llama-server on an RTX 4060 8GB VRAM, along with a thinking budget pitfall.
The successor preview model to Qwen3.6-Plus, with improved coding, knowledge, and tool-calling.
Opus 4.7 (high) took first place on the LLM Debate Benchmark.
The DeepSeek-R1 family was compressed and fine-tuned for medical use with 4-bit GPTQ+QLoRA.
Z.AI's GLM-5.1 has been added to the NVIDIA API and is offered as a free endpoint.
Qwen3.6-Max-Preview has been released as a preview with improvements in coding, knowledge, and instruction following.
It's claimed that support for temperature/top_p/top_k has effectively disappeared in Claude Opus 4.7.
Manual extended thinking has disappeared in Opus 4.7, leaving only adaptive thinking.
Qwen3.6 has been released with enhanced frontend workflows and repo-level reasoning.