Latest news, beside the leaderboard

Inception co-founder Stefano Ermon officially unveiled Mercury 3, the company's next-generation diffusion-based large language model. Delivering an astonishing 1,068 tokens per second inference speed on NVIDIA GPUs, Mercury 3 scales to 40x the size of Mercury 1 with a native 1M context window. Benchmarked on IFBench, GPQA Diamond, and SWE-Bench Atlas, it matches frontier closed baselines with standard API pricing of $0.20 / $0.70 per 1M input/output tokens (with 50% promotional pricing of $0.10 / $0.35).

Meta officially disclosed its gigawatt-scale AI supercluster Prometheus in New Albany, Ohio, housing hundreds of thousands of frontier GPUs with 85%+ pretraining goodput and 90%+ in reinforcement learning. Powered by this infrastructure, Meta's latest coding flagship Muse Wonder Full surged to #6 globally on LMSYS Code Arena: WebDev with 1726 Elo points (surpassing Claude Sonnet 5.5 and Gemini 4 Argon), lifting Meta to the #3 AI research lab worldwide.

Researchers from Amazon Web Services (AWS) and UIUC unveiled Lean4AgenticRL (arXiv:2610.22448). Tackling the limitations of conventional agentic reinforcement learning relying on sparse, binary outcome rewards, the framework grounds multi-level process rewards within a formal Lean4 verification schema. Utilizing workflow-stage-normalized GRPO, the resulting LAR-27B model reaches 77.3% on TerminalBench-2.1, outperforming base models by 9.89% and surpassing comparable open models by 5.07%.

ControlNet creator Lvmin Zhang and researchers from Stanford and CUHK unveiled MTEdit (arXiv:2610.23751). To resolve the visual degradation and cumulative error drift inherent in conventional sequential rollout for multi-turn image editing, the team introduced Packed Trajectory Canvas (PTC). MTEdit constructs a landmark dataset of 500K trajectories and 2.69M editing turns (averaging 5.38 turns, up to 8 turns), demonstrating stable instruction following and ambient identity consistency across 8 consecutive edits.

On October 19 Dream Labs came out of stealth with a $90M seed round and opened a public API for Dream-1. Dream-1 is a World Action Model (WAM) that learns to predict video and actions together. The company says it can zero-shot control bimanual parallel-gripper robots without any demonstrations from that robot, and it ran 27 unseen robots in simulation, including the 18-year-old PR2. Pricing is $3/hour for Dream-1 and $2/hour for Dream-1-flash. In an independent Robocurve evaluation, Dream-1 scored 59 on RoboDojo-RC Tier 1 versus 48 for the previous best. The company stresses it is a research beta not validated for safety-critical use.

On October 19 Inception released Mercury 3, which it calls the most capable diffusion LLM (dLLM) to date. Per the vendor, it is more than 40x the size of Mercury 1, outputs 1,068 tokens/s, expands context from 260K (Mercury 2.5) to 1M tokens, and beats GPT-6 Luna and Claude Haiku 5.5 (both medium) on GPQA Diamond, IFBench, SWE-Bench Atlas and more. List price is $0.20/M input and $0.70/M output, 50% off at launch. It is live on the Inception API, OpenRouter, Baseten and Vercel AI Gateway with an OpenAI-compatible API. All scores are vendor-reported and not yet independently reproduced.

Meta's engineering blog on October 19 gave the first detailed look at Prometheus, the New Albany, Ohio supercluster that trained the Muse family, including Muse Wonder. Phase one is complete with hundreds of thousands of GPUs, and the full build targets about 1 million. Muse Wonder training reached more than 85% goodput in pretraining and more than 90% in reinforcement learning, with a mean time between interruptions of about 47K GPU hours, since raised above 54K by firmware fixes. It runs GB200 NVL72 and GB300 NVL72 platforms, with Vera Rubin NVL72 arriving soon.

LMSYS officially announced a 14-day exclusive direct testing window for OpenAI flagship vision generative model GPT-Image-2.5 Sunburst in Direct Mode (active until Nov 2, 8 a.m. PT). Currently holding the undisputed #1 ranking across LMSYS Text-to-Image, Image Edit, and Multi-Image Edit arenas, the model can now be directly chosen from the dropdown menu before returning exclusively to anonymous Battle and Agent Mode.

Researchers from Stanford University and UC Berkeley introduced Librarian Agents (arXiv:2610.22273). To overcome the bottleneck of reactive RAG methods struggling with complex cross-document reasoning at query time, the framework uses reinforcement learning to train LLMs to perform anticipatory directed curation on messy long-context workspaces into structured synthetic filesystems. On BrowseComp+, it hits 55.86% accuracy while reducing navigation token consumption by 33.8% to 51%.

Researchers unveiled LiveMACEBench (arXiv:2610.09872), the first process-aware benchmark for persistent LLM agents operating in live, continuously evolving financial markets. Evaluating frontier models including GPT-5.4, DeepSeek-V3.2, Gemini-3.1-Pro, Qwen3-Max, and Grok-4.20 across multi-round closed loops, the study reveals that surface returns frequently mask decision vulnerabilities—such as silent compliance without audit trails or boundary targeting errors that breach safety caps.

MiniMax officially announced M3.1 Flash, its next-generation lightweight frontier model (previously dominating leaderboards as the secret model Space Bunny across 200T+ tokens). Featuring a native 1M context window, flat-rate API pricing of $0.15 / $0.60 per 1M input/output tokens ($0.015 cache reads), and 3x faster inference at 1/3 the cost of M3. It surged +106 points to 1588 on LMSYS Code Arena: WebDev (#12 among open models), with full weights set for open release on October 23.

On October 19 Ant Ling (inclusionAI) open-sourced five Ming-3.0 models under MIT, spanning speech synthesis (tts-0.5B; tts-tiny, 7.9B total/1.3B active), audio understanding (audio-flash, 126B/6.5B active; audio-tiny, 9.3B/2.7B active) and speech recognition (asr-tiny, 8.6B/2B active). Per the vendor, tts-0.5B ranks #1 among open-weight models on Artificial Analysis Controlled Voice at 1069 Elo and ties #1 overall on pronunciation robustness at 93.2%; asr-tiny averages 4.41% error across four ASR categories. Weights are on Hugging Face and ModelScope, with day-0 SGLang and vLLM support.

In small language models, vocabulary embedding tables often account for over 60% of total parameters, posing severe bottlenecks for edge deployment and ternary quantization. A new study (arXiv:2610.22428) replaces trainable embeddings with a single frozen 16 MB table of 512-bit binary codes derived via PCA and sign binarization. On Gemma 3 270M, trainable parameters collapse from 268.1M to 101.0M (-62.3%) while matching or exceeding original performance across HellaSwag (30.4%), LAMBADA (23.2%), and PIQA (61.5%).

Generating usable, physically coherent 3D indoor scenes from text prompts requires specialized design and aesthetic critique. Researchers from HKU and NTU open-sourced Scene-Designer (arXiv:2610.23121, 29 upvotes on Hugging Face). The agentic framework orchestrates a Designer Agent and a Critique Agent through complementary reinforcement loops, surpassing prior systems like Holodeck and Chat-3D in layout functionality, spatial plausibility, and user preference scores.

SemiAnalysis published its major forward-looking industry analysis 'Datacenters 2027: The Year of Acceleration' (X: 2112069964131713517). The report reveals an unprecedented +50GW of capacity added in a single year, 60GW under construction, and over $1.3T in signed lease contracts year-to-date. As massive power and real-estate infrastructure projects come online by 2027, the primary bottleneck in the AI supply chain is projected to pivot from grid power back to leading-edge accelerator silicon.

Multimodal LLMs excel at holistic image captioning but struggle with pixel-level spatial grounding due to lack of spatial supervision. Xiaomi AI Lab unveiled Xiaomi-Lens-0 (arXiv:2610.23279), a 4B MLLM based on Qwen3.5 that natively predicts pixel masks. A lightweight decoder integrates patch embeddings, visual latents, and mask-token activations, driving RefCOCO comprehension (REC [email protected]: 94.6% / 92.1% / 92.5%) and segmentation (RES cIoU: 83.4% / 79.9% / 81.1%) past Florence-2 and LISA-7B while preserving a 69.7% MME-RW score.

Viggle released JST-2 on October 19. Instead of learning in pixel or video space, it represents the world as persistent "atoms" in 3D space (Physics-Native Tokens) and uses one architecture to generate scenes, characters and motion. Viggle says a 3D scene takes about one second on an NVIDIA B200 at under $0.01 of inference, with up to 64 input views kept as spatial context, and characters animate without a separate rigging step. It is free to try in the PINOC browser app and is also available via API and MCP.

Modal said in an October 19 blog post that it will add NVIDIA's Arm-based Vera CPU to power Modal Sandboxes, its first step beyond an all-x86 fleet. Using Agentless Agent Bench, a 50-task suite built from Terminal-Bench, Modal measured up to 2x the task throughput of a comparable x86 GNR machine under heavy concurrency, with per-task latency equal or lower (vendor figures). On an idle host running tasks serially, Vera's edge was small. The capacity targets background coding agents and reinforcement learning rollouts.

MiniMax confirmed that Space Bunny, the anonymous model that led usage rankings on OpenRouter and elsewhere, was an early version of M3.1 Flash, and launched it on its API on October 19. Pay-as-you-go pricing is $0.15 input, $0.60 output and $0.015 cache read per million tokens, with a 1M context, text, image and video input, and an effort setting for thinking depth. Per MiniMax, it costs one third of M3 and runs 3x faster, scoring 73.7% on DeepSWE and 62.2% on Automation-Bench; weights are due October 23.

On October 19 Amazon announced TorchNeuron, which uses PyTorch's standard PrivateUse1 device-backend mechanism to register Trainium as a native "neuron" device, with eager mode, torch.compile and native distributed APIs such as FSDP, DTensor and DDP on Trn2 and Trn3. Amazon says existing PyTorch code runs with a single device configuration change, and it will be open-sourced under Apache 2.0 at PyTorch Conference North America on October 20. The official docs still list it as a closed beta for now.