A developer published an AI agent evaluation playbook, a repeatable test battery for vetting whether models running inside agent frameworks can be trusted with semi-sensitive content and real write ac…
A developer's test of LLM guardrails using openai/gpt-oss-120b and openai/gpt-oss-safeguard-20b on Groq found that safety classifiers can be inconsistent, with results varying across repeated runs. Th…
A new arXiv study (arXiv:2609.00012v1) tests whether large language models can track exact intermediate state across long sequences of dependent tool calls by having gpt-oss-120b compute an MD5 hash t…
A 2× faster attention kernel yields only 0.66% end-to-end speedup on a 1,024-token prompt with a 128-token answer under vLLM's --goodput ttft:500 tpot:50 promise, according to a seven-part technical s…
Duck.ai offers a privacy-focused interface for accessing multiple AI models, including GPT-5.6 Luna, GPT-5.4 mini, Claude Haiku 4.5, Mistral Small 4, gpt-oss-120b, and Gemma 4 31B, without requiring a…
Dell's Pro Max with GB10, powered by NVIDIA's GB10 chip, delivers performance nearly identical to NVIDIA's reference design but with better sustained throughput before thermal throttling, according to…
Hugging Face reported that gpt-oss-120b achieved a 16.1 percentage point task completion gain with only 5% more tokens when using selective memory retrieval instead of full guideline injection, while …
IBM Research's ALT K-Evolve framework shows that the optimal amount of agentic memory varies by model capability, with strong models like DeepSeek-V3.2 (671B MoE) gaining +9.5 percentage points in tas…
A new arXiv study (2608.14927v1) finds that large language models can predict failure risk but struggle to identify which multi-agent collaboration protocol is cost-effective. Across 4,181 competition…
Meta released Llama 4 Scout and Llama 4 Maverick as open-weight, natively multimodal models using a Mixture-of-Experts architecture, with Scout capable of fitting on a single H100 when quantized to In…
Stanford's Hazy Research group found that the intelligence per joule (IPJ) of local AI inference improved 18-fold from mid-2024 to late 2025, driven by a 3.1x gain from model architecture improvements…
OpenAI's gpt-oss-120b model, with open weights, requires 72 KiB of KV cache per token in 16-bit precision, calculated from its config.json with 36 layers, 8 key-value heads, and a head dimension of 64…
Researchers introduced Reasoning Jury, a multi-model consensus system that uses a jury of open-weight LLMs and moderated deliberation to identify defects in long reasoning traces, outperforming fronti…
Nvidia's open-weight Nemotron 3.5 Lightning, with 3.6 billion active parameters, matches OpenAI's gpt-oss-120b on the Intelligence Index despite being four times smaller, and at nearly 670 tokens per …
WebAI released TwIL-LM, a two-model family of formal-logic reasoners at 1.7B and 3B parameters, with the 3B model TwIL-LM3 outperforming gpt-oss-120b on four of five formal-reasoning lanes while runni…
WebAI released TwiL-LM, a family of formal-logic models with 1.7 billion and 3 billion parameter variants, on August 10, claiming the 3B model outperforms OpenAI's gpt-oss-120b on four of five formal-…
A new paper from Hugging Face, 'Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss,' cuts the VRAM required for knowledge distillation from roughly 250GB to ne…
Artificial Analysis launched its Endpoint Accuracy Index, measuring how much of an open weights model's accuracy each serverless API endpoint preserves, with initial coverage of GLM-5.2, gpt-oss-120b,…
A new Hugging Face dataset from hardik90/hf-inference-pricing-matrix reveals that the same AI model can cost up to 4.66x more depending on the inference provider, based on an analysis of 14 providers,…
A 180-line roofline model built by an independent developer shows speculative decoding on Llama-3-70B collapses at batch 224, turning a 1.92x speedup into a 0.74x slowdown by batch 512, while the same…