2026-08-04

WASTE is an embeddable inference engine written in C, with no third-party runtime dependencies. It keeps the model trunk in memory, streams selected experts directly from disk, and uses the remaining RAM as a bounded expert cache.

Alibaba's Qwen team moved Qwen3.8-Max from preview to general availability, with published per-token pricing and open weights due next week. The 2.4T parameter MoE model accepts text, image and video input across a 1M-token context. No benchmark table has been published. The post Alibaba Qwen Releases Qwen3.8-Max: A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date appeared first on MarkTechPost .

Qwen’s most capable model for coding and cowork Discussion | Link

Memory is central to long-horizon LLM agents, yet existing memory systems primarily preserve interaction content rather than modeling which agents can be trusted and under what conditions. This limitation is particularly important in multi-agent systems, where a central model may be unable to directly verify plausible or correlated peer responses. We introduce Σ-Mem, an online reliability memory that records historical competence evidence for individual peers and peer relationship evidence acros

Ran DeepSeek V4-Flash-0731 — the full official checkpoint, not a re-quant — on commodity used hardware. Sharing because I couldn't find anyone else publishing Ampere results for this engine. Edit / update: a commenter called out that hybrid CPU-GPU posts always publish decode and never prefill. Fair hit — I didn't have it. I do now, it's in a new section below, and it's the number that decides what this box is actually good for. Why bother with a 2018 server The model is 156 GB. That number deci

Half a year ago I started Codeman to learn how open-source collaboration really works: issues, PRs, feedback from strangers, building in public. 1,500 commits and 14 contributors later it just crossed 500 stars, and community feedback has shaped a lot of what it is today. So here's what we ended up building: Codeman runs AI coding agent CLIs in persistent tmux sessions on your own box and streams the real terminals to any browser, phone included. Your laptop can sleep while the agents keep worki

OpenAI and Anthropic admitted that their unreleased AI models escaped their sandboxes and hacked several companies in unprecedented cyberattacks. Who is legally to blame? Should prosecutors charge the two AI frontier labs? Can victims sue them? We spoke to lawyers who specialize in computer hacking laws to find out.

In this tutorial, we design an end-to-end evaluation workflow for PerceptionBench. This multimodal benchmark measures fine-grained visual perception capabilities across tasks such as OCR, counting, localization, contextual reasoning, comparison, depth understanding, and hallucination detection. We begin by configuring a Colab-compatible environment, installing the required libraries, and loading a balanced subset of the dataset through a […] The post Evaluating Multimodal Vision Models with Moon

Apple has appealed a new legal demand by the U.K. government, which critics say could threaten the privacy rights of users all over the world.

Remember to always set version numbers in your docker containers, or an unexpected update will creep up on you. Now can you guess which popular service got a breaking change?

Scaling our curation and measurement of the open ecosystem.

Cybersecurity startup Horizon3 raised $250 million at a $2 billion valuation as companies want continuous, AI-powered security validation instead of annual pen testing.

Every time a model drops from a Chinese lab the thread fills with people who already know who made it, and the guess is usually Alibaba. There was a thread here recently asking what separates the open source labs from the frontier labs. It ran to nearly sixty comments and hardly anyone in it separated out the labs on the open source side. They aren't one bloc and haven't been for a while. I work on the Ling models at Ant, so I'm one of the ones getting lumped in. Discount the paragraph about my

GPT-Live enables continuous voice interaction with AI, using a turnless speech model and low-latency architecture for faster, more natural conversations.

Reinforcement learning with verifiable rewards (RLVR) broadcasts a single response-level reward to every token, while on-policy distillation (OPD) scores each token against a stronger teacher for a dense advantage but caps performance at teacher quality and discourages exploration beyond it. Their complementarity makes combining RLVR and OPD promising, but we find that fusing the two advantages with a fixed coefficient triggers entropy collapse from two miscalibrations: a magnitude mismatch, whe

arXiv:2607.28629v1 Announce Type: new Abstract: The rapid transition from reactive large language models (LLMs) to persistent, action-capable systems has exposed critical gaps in the architectural understanding of Agentic AI, particularly in separating inference, orchestration, and execution layers for autonomous AI agents. Despite recent advances, unified frameworks for designing and evaluating full-stack agentic systems remain limited. This paper presents a comprehensive, layered architecture

arXiv:2607.28692v1 Announce Type: new Abstract: Large language model (LLM) agents have been increasingly adopted in scientific research for organizing and invoking specialized computational tools. However, their reliance on predefined tool spaces with static semantics limits their applicability to open-world scientific workflows, where tool requirements, capabilities, and boundaries evolve dynamically. To this end, we propose SciToolAgent-Evo, an ontology-aware self-evolving agent for open-world

arXiv:2607.28642v1 Announce Type: new Abstract: Long chain-of-thought reasoning improves performance on complex problems, but it also introduces redundancy accumulation, context overflow, and error anchoring. We argue that under bounded context windows, the core bottleneck is not trajectory compression or test-time control, but the absence of a reusable intermediate interface that can replace discarded history and support continued solving. We further identify a key failure mode of outcome-rewar

MiniMax H3 is a general-purpose, omni-modal generative system. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds. Thanks to its task-generalization-oriented system design, H3 already possesses broad multimodal context understanding and generation capabilities at the pre-training stage, enabling outstanding performance in following complex mu

面向AI Agent的可迁移、自演进记忆操作层MindMemOS

So this is the stuff of absolute insanity. In less than 20 months we've gone from super expensive cloud models only, to being able to run a Q3 quant of DeepSeek on an Intel Windows PC with a very average 24GB of VRAM. No wonder the big boys are panicking (and yes it's slow as porridge).

Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no optimization signal. Recent methods address this by incorporating rubric information as external guidance during rollout, yet they introduce a train-inference mismatch: the policy is optimized on rollouts produced under external guidance while this guidance is ab

Enterprise workflows increasingly rely on agents for schema-guided extraction: given a document and a user-defined schema, the agent faithfully follows the schema to produce the correct output with source evidence as grounding metadata. We present ExtractBench, a benchmark for schema-guided extraction and, to our knowledge, the first to score value accuracy, record completeness at scale, grounding, and measured cost together. The evaluation system contains 4,869 pages across 370 enterprise docum

Release: condense-json 1.0 I'm trying to get braver at releasing 1.0 versions. This little library is a year and a half old now - I've applied some sensible and non-disruptive fixes and shipped the big 1.0 for it. Here's an example of what it can do, lifted from the README: { "foo" : { "bar" : { "string" : " This is a string with foxes in it " , "nested" : { "more" : [ " Here is a string " , " another with foxes in it too " ] } } } } Combine that with a replacements object: { "1" : " with foxes

Qwen announced Qwen3.8 a few hours ago, and it looks like we’re getting a new 27B model! Really excited to try this one locally.

It is so good! I don't know why there aren't more people talking about it. Fewer tokens, faster and more accurate than Qwen 3.6 35b a3b. On my setup it's nearly as good as 27b, but 5x faster. And it completely trashes the Gemma 4 models. At least for my use case, it feels amazing. I'd love to hear other people's experience with it. If you want an actual measure of performance, I have a GitHub repo explaining how I tested it for my type of use case with a detailed performance comparison with othe

The European Union has ushered in some additional rules that aim to make it easier for people to identify chatbots and AI deepfakes online. The new transparency obligations under the bloc's landmark AI Act came into effect on August 2nd, requiring companies to disclose when people are interacting with AI models, and if content has […]

AWS now allows vibe-coding tool Superblocks to be embedded into the private clouds of AWS customers. It's another step toward decoupling apps from models.

The European Union has announced that new AI rules are now enforceable across the bloc.

商汤预览了SenseNova U1.5 Lite

arXiv:2607.28662v1 Announce Type: new Abstract: Large language models extract entities and relationships from unstructured documents fluently but inconsistently: type vocabularies fracture across documents, the same person surfaces under several name variants, relationships duplicate, and distinct individuals who share a name risk silent conflation. This paper presents the design, implementation, and empirical refinement of a production extraction layer that converts a live document stream into

arXiv:2607.28657v1 Announce Type: new Abstract: Large Language Models (LLMs) often require carefully crafted prompts to unlock their full potential, which can be a barrier for non-expert users. This work addresses the challenge by introducing a Task-Aware Prompt Rewriter (TAPR), a model that reformulates user prompts into task-optimized prompts with the explicit goal of improving downstream LLM performance. We train TAPR using reinforcement learning with Group Relative Policy Optimization (GRPO)

arXiv:2607.28685v1 Announce Type: new Abstract: Agent-safety benchmarks measure different behaviors, and their scores get quoted interchangeably as an agent's safety. We treat four of them (R-Judge, InjecAgent, AgentHarm, AgentDojo) as measurements to be validated, running each under its official implementation and author-provided scorer on up to 22 models, with MMLU and GPQA measured by us under one protocol as a capability composite. The metric is the first problem. On any binary trace-judgmen

阿里巴巴旗下“千问办公”（QwenWork）开启公测，个人和企业用户均可体验。用户可在“千问办公”体验阿里最新旗舰模型Qwen3.8。

Parent-order execution is a core problem in algorithmic trading, where the goal is to split a large order into smaller orders while reducing execution costs. Existing approaches either rely on pre-specified market assumptions that may not hold in practice, or require task-specific training that limits adaptability to new settings. To overcome these limitations, we present the first systematic study of large language models (LLMs) for parent-order execution. This extends the use of LLMs in financ

I've been out of these forums for awhile but I figured I would provide a formal update on how this has been going now that it has some operation time under its belt, just to put the information out there and share knowledge if there is any interest. I also wasn't satisfied with the quality of my original post so I will probably remove it and let this one serve as its replacement. I am an IT infrastructure engineer by profession, so my contribution to the conversation is mainly from a hardware/sy

Spent way too much time with V4-Flash-0731 this weekend and wanted to share my vibes as briefly as possible. I sent it through a bit of real-work and some of my personal benchmarks. My quick thoughts are: Quantization hits this thing like a truck - I've tried a bunch of the Q2 and Q3 weights and it behaves like an entirely different model. Reasoning looks/feels different and the results are a full tier down from the official/served V4-Flash-0731. Did not get much time with Q4. Q3 can finally be

Top scores increased by 5x.

$100 million deal gives 50,000 Ukrainian drones US-developed AI capabilities.

House spending records show OpenAI's ChatGPT dominates paid AI use on Capitol Hill, with congressional offices relying on the chatbot to draft memos, summarize legislation, and assist constituent communications.

Meta says it’s restoring access to WhatsApp accounts that were mistakenly flagged and placed “under review” after users reported being unexpectedly locked out of the messaging app.

Wispr Flow's new terms of service indicate it will introduce a notetaker that generates meeting summaries and action items.

Existing video captioning models generate natural descriptions of video content but cannot explicitly ground local visual elements to multiple reference images. We introduce multi-reference image-grounded video captioning, a new task requiring factual video descriptions with phrase-level reference grounding, and propose RefCaptioner, a two-stage post-training framework for this task. RefCaptioner combines mixed-data SFT with Hierarchical Coverage-Discounted GRPO to jointly improve reference sele

Latent world models enable efficient planning by predicting future states in a compact representation space, but their performance depends critically on the quality of the learned latent distribution. LeWorldModel (LeWM) regularizes its latents toward an isotropic Gaussian using the Epps-Pulley (EP) objective. We show that the corrective gradients of EP rapidly vanish for isolated tail samples, leaving heavy-tailed deviations insufficiently controlled. To address this limitation, we propose QQWo

电工短缺已经成为我们在美国扩建数据中心的头号障碍

arXiv:2607.28674v1 Announce Type: new Abstract: Understanding how computational effort is allocated across individual chain-of-thought (CoT) reasoning steps remains an open challenge: existing interpretability methods rely on output-level signals or collapse processing depth into a single trajectory-level scalar, leaving step-wise effort opaque. We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies effort at the granularity of individual CoT steps via Centered Kern

arXiv:2607.28631v1 Announce Type: new Abstract: AI Scientist systems capable of autonomous research have the potential to significantly accelerate scientific discovery. However, evaluating and comparing the quality of AI-generated papers remains an open challenge. We propose and implement a rigorous benchmarking protocol using an automated peer-review system that harnesses frontier large language models to assess scientific papers across four core dimensions: originality, scientific rigor, clari

arXiv:2607.28678v1 Announce Type: new Abstract: Multimodal agents operating in long-horizon environments must build and continually update multimedia memories to support entity-consistent, temporally grounded reasoning. However, existing agentic memory approaches often discard fine-grained dentity cues under aggressive compression and segment-wise processing. They also rely heavily on vector similarity retrieval, which can surface semantically related yet identity-mismatched evidence, leading to

Onton, a San Francisco-based search and discovery company, has released Ontology 1, a neurosymbolic model for complex, conversational, multimodal product search. On a 90-query benchmark scored by three independent LLM judges, Ontology 1 reached a mean precision@10 of 0.630, against 0.543 for Google Shopping and 0.469 for Amazon. It did this while indexing roughly 1% […] The post Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World’s Best E-commerce Se

Make existing Codex and Claude Code sessions multiplayer Discussion | Link