Sep 30, 2026 | Gemini 4 Argon delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. Today, we’re announcing our new frontier model, Gemini 4 Argon, which is rolling out to a set of trusted cyber defenders through our Fairwind Program.
Liquid AI released experimental draft model LFM2.5-VL-DSpark for vision-language model LFM2.5-VL-3B. Speculative decoding speeds it up without changing output quality: decoding improves by up to 3.13x on devices and 2.66x on H100, with end-to-end gains of up to 2.62x and 2.27x respectively.
NVIDIA submitted preview results for Vera Rubin NVL72 to MLPerf Inference v6.1 for the first time, achieving throughput up to 3.7 times that of GB300 NVL72 on Qwen3-VL and up to 2.5 times on DeepSeek-R1.
Google Research proposed Retrieve-for-Train, using offline reinforcement learning to find reward-aligned query fan-out and compile it into supervision, then distilling it into a 53.9M-parameter diffusion retriever. At inference, it performs one non-autoregressive query fan-out without generating CoT reasoning tokens.
NVIDIA announced several advances for Vera Rubin and the DSX platform at AI Infra Summit, focusing on energy-efficiency improvements in token throughput per megawatt.
OpenAI shared an AI-generated solution to the Navier–Stokes Millennium Prize Problem, including a written explanation and a formal Lean proof.
A Hugging Face blogger reproduced Surya Narreddi's idea of having a language model paint watercolours using TRL and OpenEnv. The model paints through around 150 lines of JavaScript with p5.brush, with datasets, RL environments, training scripts and models all open-source.
Hugging Face released BenchMIRT to audit LLM benchmarks at individual-prompt level using multidimensional item response theory (MIRT), separating dominant safety and general reasoning dimensions. Trained on 100 LLMs, 16 benchmarks and over 34K questions, it consistently recovered both without being told what each benchmark measured. Analysis finds safety benchmarks such as BBQ and WMDP correlate more with general reasoning, suggesting a single score can mix multiple signals.
The IBM Granite team released Granite 4.2, its first dense, decoder-only reasoning model family, in 3B, 8B and 30B sizes, all open-source under Apache 2.0.
Multiverse Computing published a paper proposing Quantization-Aware Healing (QAH). After compressing GPT-OSS 120B to 60B parameters and quantising it to MXFP4, the method distils directly from the original uncompressed model rather than a reconstructed bfloat16 checkpoint.
Liquid AI released DSpark draft model checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. Speculative decoding accelerates decoding without changing output quality, increasing throughput by up to 3.18 times on GPUs and 2.87 times on devices.
Microsoft Research introduced MindTopo to evaluate multimodal models' reasoning and planning over connectivity, enclosure, order, separation and knotting. Current models perform much better at static-image recognition than interactive planning, with failures mainly in planning rather than perception and overall performance far below humans. Image and video generation helps only when preserving structural relations in a single frame; multi-step operations often change topology or violate constraints.
Hugging Face's ALTK-Evolve and ACE both provide agent memory without weight updates, but differ in delivery: ACE injects the full playbook at every step, while ALTK-Evolve retrieves a small number of strongly supported guidelines for each task.
Google Research released the Chain-of-Evidence (CoE) verifiability framework, implemented in a Science One Framework prototype, with automated CoE Audit metrics.
Berkeley AI Research and IBM Research extended the K-Search evolutionary kernel search framework to MLX, transferring existing CUDA kernel knowledge to Apple Silicon through a structured CUDA-to-MLX translation layer.
Berkeley AI Research proposed ABBEL, isolating and supervising LLM summaries as natural-language belief states and using autoencoder-inspired reconstructive belief scoring as an auxiliary RL task. On CollabBench, ABBEL-rec-BG narrowed the performance gap to the full-context model by around 50%, reduced training steps from 100 to 50 and used a shorter peak token length.
Thinking Machines released Inkling on Hugging Face, a multimodal MoE model with around one trillion parameters, a one-million-token context and native image, text and audio inputs, alongside Inkling-Small with 276 billion total and 12 billion active parameters.
The third Hugging Face PyTorch profiling instalment traces attention implementation with a profiler. Comparing naive attention with in-place masking shows replacing masked_fill with masked_fill_ removes one GPU Memcpy kernel per forward pass.
Mistral released Leanstral 1.5 under Apache-2.0, with 119B total and 6B active parameters and major formal verification improvements. It achieves 100% on miniF2F, solves 587 of 672 PutnamBench problems, and sets current best results of 87% on FATE-H and 34% on FATE-X.
Google Research published an architecture attaching multi-token prediction (MTP) heads to frozen Gemini Nano v3 models, accelerating on-device inference without changing backbone weights. It has shipped with the Pixel 9 and 10 series.
A Google Research paper at COLM 2026 finds reasoning traces unlock factual knowledge LLMs otherwise cannot recall, even for simple single-hop questions. Tests on Gemini-2.5 Flash and Pro and Qwen3-32B identify two mechanisms: extra tokens act as a 'computation buffer', and 'factual priming' produces related facts to semantically prepare the correct answer. Self-generated intermediate facts can also introduce hallucination risks.
Google DeepMind released experimental open-source model DiffusionGemma, generating text blocks in parallel through text diffusion and achieving up to 4 times faster inference on dedicated GPUs.
Berkeley AI Research reviews parallel reasoning, focusing on models deciding when to decompose and parallelise independent subtasks, how many threads to generate and how to coordinate. Existing approaches including Self-consistency, Best-of-N, Tree of Thoughts, MCTS, ParaThinker, GroupThink and Hogwild! Inference mostly impose parallel structures externally rather than teaching adaptive behaviour.
Berkeley AI Research proposed GRASP, a gradient-based planner for learned world models. It lifts trajectories into virtual states for parallel optimisation across time, injects randomness directly into state iterations for exploration and reshapes gradients to give actions clear signals. Avoiding fragile state-input gradients in high-dimensional visual models makes long-horizon planning more practical and robust.
Google proposed Simula, a reasoning-first synthetic data framework reframing dataset construction as mechanism design. It builds datasets from scratch without seed data through global diversification, local diversification, complexification and dual-critic quality checks.
Google Research released TurboQuant, compressing the KV cache to three bits without loss of model accuracy or training and fine-tuning, alongside the QJL and PolarQuant methods.
Mistral AI released Mistral Small 4, the next major version in the series and its first model to combine Magistral reasoning, Pixtral multimodal capabilities and Devstral agentic coding in one model, under Apache 2.0.
Mistral AI released Leanstral, the first open-source coding agent for Lean 4, using a sparse architecture with 6B active parameters. Its weights are available under Apache 2.0, and it is integrated into Mistral vibe and the free labs-leanstral-2603 API endpoint.
Berkeley AI Research proposed SPEX and ProxySPEX to identify key interactions driving LLM outputs at scale in feature, data and model-component attribution. SPEX turns interaction search into sparse recovery using sparsity and low order; ProxySPEX exploits hierarchy to match SPEX with roughly ten times fewer ablations.
Berkeley AI Research proposed a divide-and-conquer off-policy reinforcement learning algorithm without temporal-difference (TD) learning. It reduces Bellman recursions from linear to logarithmic counts, scaling to long-horizon tasks.
Mistral added features to Le Chat including preview Deep Research, Voxtral-powered voice mode, multilingual reasoning with Magistral, Projects for organising conversations and image editing with Black Forest Labs.
Mistral AI released its first reasoning model, Magistral, comprising the 24B open-source Magistral Small and enterprise-focused Magistral Medium.