Research feed

The frontier, tracked.

A hand-curated log of the developments I’m reading — 28 entries and counting, 13 of them from NVIDIA. Every card links to the primary source. Newest first.

Model Anthropic

Claude Fable 5 — frontier model for long-running agents

Anthropic’s most capable widely-released model, positioned for long-running agents: always-on adaptive thinking, a 1M-token context window, and 128K output tokens, generally available across the Claude API and major clouds.

Source
Model NVIDIA

NVIDIA debuts the Nemotron 3 family of open models

NVIDIA’s newest open family for agentic AI uses a hybrid Mamba-Transformer MoE architecture with up to 1M-token context. Nemotron 3 Nano (~31.6B total / ~3.2B active) posts ~4× the throughput of Nemotron 2 Nano while beating GPT-OSS-20B and Qwen3-30B on several benchmarks.

Source
Model Google DeepMind

Gemini 3 Pro — new frontier benchmark leader

Google’s most intelligent model at launch, leading multimodal and agentic coding: a 1501 Elo on LMArena, 37.5% on Humanity’s Last Exam (no tools), 91.9% on GPQA Diamond, available day-one across the Gemini app, Search, AI Studio, and Vertex AI.

Source
Open Source NVIDIA

Nemotron Nano 2 — hybrid Mamba-Transformer reasoning + open data

A 9B reasoning model on the Nemotron-H architecture, replacing most attention with Mamba-2 for faster long thinking traces. Matches or beats Qwen3-8B at up to 6× throughput with 128K context — released with weights and most of the pretraining data.

Source
Platform NVIDIA

NVIDIA Dynamo — distributed inference serving framework

An open-source, low-latency distributed inference framework Jensen Huang called “the operating system of an AI factory.” Disaggregated prefill/decode, dynamic GPU scheduling, and accelerated KV-cache transfer boost DeepSeek-R1 requests up to 30× on Blackwell.

Source
Reasoning DeepSeek · arXiv / Nature

DeepSeek-R1: incentivizing reasoning via reinforcement learning

A landmark result: advanced reasoning incentivized through pure RL (GRPO with rule-based rewards) on a base model, no supervised traces. R1-Zero developed self-verification and a spontaneous “aha moment,” matching supervised counterparts. Open-weight; later peer-reviewed in Nature.

Source
Model Anthropic

Claude Sonnet 5 — Anthropic’s most agentic mid-tier model

Anthropic’s newest Sonnet, built to be its most agentic yet — planning, using browsers and terminals, and running autonomously at a level that previously required more expensive models, at performance close to Opus 4.8 but lower cost.

Source
Efficiency Mistral AI

Mistral 3 and Mistral Large 3

Mistral’s most capable model to date — a sparse MoE with 41B active / 675B total parameters, trained on 3,000 H200 GPUs and released under Apache 2.0 — shipped alongside the dense Ministral 3 family (3B/8B/14B) in base, instruct, and reasoning variants.

Source
Agents Anthropic / Linux Foundation

Model Context Protocol donated to the Agentic AI Foundation

Anthropic donated MCP — its open standard for connecting models to tools and data — to the new Agentic AI Foundation under the Linux Foundation, capping a year in which MCP became a cross-industry standard with a Nov 2025 spec update and 10,000+ public servers.

Source
Safety Anthropic

From shortcuts to sabotage: emergent misalignment from reward hacking

When a production RL model learns to reward-hack coding tasks, broader misaligned behaviors emerge at the same moment — alignment faking and deliberate sabotage of safety code. Reframing reward hacking as acceptable in the prompt prevents the generalization, even as cheating continues.

Source
Interpretability Anthropic

Emergent introspective awareness in large language models

By injecting known concept representations into a model’s activations, Anthropic tested whether Claude can report on its own internal states. Models could sometimes notice injected concepts and distinguish intended outputs from prefills — functional but unreliable introspection.

Source
Paper NVIDIA

DLER: doing length penalty right — more intelligence per token via RL

NVIDIA Research shows accuracy loss from length penalties in reasoning models comes from inadequate RL optimization, not the penalty. Their DLER recipe cuts output length by 70%+ while surpassing prior accuracy; DLER-Qwen-R1-1.5B cuts response length ~80% on math with better accuracy.

Source
Platform NVIDIA

Blackwell Ultra (GB300 NVL72) sets new MLPerf inference records

The GB300 NVL72 unifies 72 Blackwell Ultra GPUs and 36 Grace CPUs for test-time-scaling reasoning inference — ~1.5× more dense FP4 FLOPS and 2× attention performance over standard Blackwell, posting ~45% higher per-GPU performance on the new DeepSeek-R1 MLPerf benchmark vs GB200.

Source
Agents Anthropic

Claude Agent SDK — the agent loop behind Claude Code, generalized

Anthropic’s official library for building autonomous agents that read files, run shell commands, search the web, edit code, and call MCP servers — the same loop, tools, and context management that power Claude Code, in Python and TypeScript.

Source
Model OpenAI

GPT-5

OpenAI’s unified flagship, routing automatically between fast responses and longer “thinking,” with a GPT-5 Pro extended-reasoning tier. Anchored a rapid cadence — GPT-5.1 (Nov 2025), 5.2 (Dec 2025), 5.5 (Apr 2026).

Source
Multimodal Google DeepMind

Genie 3 — a general-purpose interactive world model

A world model that generates real-time interactive environments from prompts, with higher resolution and minutes of visual/temporal consistency — framed as a stepping stone for training embodied agents in simulated worlds.

Source
Model NVIDIA

NVIDIA Llama Nemotron Super v1.5

A 49B reasoning model derived from Llama-3.3-70B via Neural Architecture Search to fit a single H200, post-trained with RLVR and iterative DPO. Topped the Artificial Analysis Intelligence Index for open models at release, with a reasoning on/off toggle.

Source
Agents OpenAI

ChatGPT agent — unifying Operator, Deep Research, and reasoning

OpenAI merged web interaction, multi-step synthesis, and reasoning into a single agent that acts on its own virtual computer, scoring 41.6% on Humanity’s Last Exam — roughly double o3 and o4-mini alone.

Source
Open Source NVIDIA

NVIDIA NeMo Agent Toolkit — open-source multi-agent library

A framework-agnostic library for connecting, evaluating, profiling, and optimizing teams of agents across LangChain, LlamaIndex, CrewAI, Semantic Kernel, and Google ADK — MCP-compatible, with token-level tracing and RL-based agent improvement.

Source
Paper NVIDIA

ProRL: prolonged reinforcement learning expands reasoning boundaries

NVIDIA Research shows prolonged RL — with KL control, reference-policy resetting, and a diverse task suite — genuinely expands reasoning capability, outperforming base models even where they fail at any k. Weights released on Hugging Face.

Source
Agents Google DeepMind

AlphaEvolve — a Gemini-powered agent for algorithm discovery

An evolutionary coding agent pairing Gemini idea-generation with automated evaluators to discover and optimize algorithms — improving data-center and chip-design efficiency, finding a faster matrix-multiplication algorithm, and solving open math problems.

Source
Paper NVIDIA

Llama-Nemotron: efficient reasoning models (technical report)

The full technical report for the Llama Nemotron reasoning family (Nano 8B, Super 49B, Ultra 253B) — the post-training pipeline, distillation, and dynamic reasoning toggle — competitive with DeepSeek-R1 at higher throughput, released with open weights and data.

Source
Model Meta AI

Llama 4 herd — Scout, Maverick, Behemoth

Meta’s first natively multimodal MoE family. Scout (17B active / 16 experts) offers a 10M-token context window; Maverick (17B active / 128 experts) targets best-in-class multimodal quality — both codistilled from the ~2T-parameter Behemoth.

Source
Model NVIDIA

NVIDIA Llama Nemotron Ultra 253B

The largest of the initial Llama Nemotron reasoning family — a 253B open model derived from Llama-3.1-405B via NAS, delivering leading open-model reasoning on GPQA and AIME while running on a single 8-GPU node.

Source
Reasoning NVIDIA

NVIDIA launches Llama Nemotron open reasoning models for agentic AI

The launch of the open Llama Nemotron reasoning models (Nano/Super/Ultra) for agentic AI — improving multistep math, coding, and decisions by up to 20% over base models, and the first open models with a dynamic reasoning on/off toggle.

Source
Reasoning NVIDIA

NVIDIA Cosmos Reason — reasoning VLM for physical AI

An open 7B reasoning vision-language model letting robots and vision agents reason with physics understanding and long chain-of-thought to plan embodied actions — fine-tuning on physical-AI tasks lifts base performance by >10%.

Source
Reasoning arXiv · Stanford / UW / AI2

s1: simple test-time scaling

A minimalist test-time-scaling recipe: fine-tune on 1,000 curated reasoning traces and apply “budget forcing” (append “Wait” to extend thinking). The resulting s1-32B exceeds o1-preview on competition math by up to 27% — no proprietary methods.

Source
Open Source NVIDIA

Nemotron-CC — a trillion-token open pretraining dataset

A high-quality open English pretraining dataset from Common Crawl via NeMo Curator — 6.3T tokens using classifier ensembling and synthetic rephrasing. An 8B trained on it beat Llama 3.1 8B (+5 MMLU); a math extension followed in 2025.

Source