2026-09-08

“在两个模型之间找到数学上的共同基础其实非常困难”

OpenBMB's MiniCPM5-2B scores 15 on the Artificial Analysis Intelligence Index v4.2, the highest of any open weights model at 4B parameters or below Hugging Face: GitHub: github.com/OpenBMB/MiniCPM

TL;DR My TAK quant of Qwen 3.8 27b scored 82.81% on reasoning, comparted with 77.34% for the byte matched Unsloth UD IQ2_S and 83.59% for BF16. Edit: Some of you have tried coding with this reasoning-specialized quant and encountered repetition loops. Coding is outside its intended domain, but I’ll investigate to reproduce and characterize the failure. Over the last few months, I've been exploring task aware quantization. I've now turned that work into a clean, repeatable pipeline under the reas

When multiple companies are behind one project, who bears responsibility for problems?

They send all typed into editor input to codepen.dev almost immediately (you would see in 1-2 sec after you typed your secret that it appears in respective Network/Response tab) even before one saved it. I tested this with a unique marker: after typing it into index.html, CodePen ran a build with "save:false", and the marker then appeared verbatim in the HTML served from the generated "*.codepen.dev preview". Thus, if you ever entered some secrets in there by mistake consider them compromized ev

Run multiple local coding agents on your machine Discussion | Link

Curious about what people are preferring, if you have the hardware. I have m3 Max 96gb and both run, and largely feel identical, but prefill on qwen 27b is faster. Is there anything / anyone working on anything to improve pp with mlx? Branching question: is anyone working on a harness that works with no reasoning? This interests me ever since Jetbrains shared that they're using 3.6 with reasoning off entirely: Feel like there must be something neat with using one model to orchestrate, with reaso

Hi everyone! We just released Rustuna (GitHub: ), a high-speed, memory-efficient implementation of Optuna built in Rust. Optuna-Compatible Design: Keeps the familiar API and concept of Optuna. Zero Python Dependencies: Mitigating the risk of supply chain attacks. Lower Memory Footprint: Optimized memory management natively in Rust. For details, please check out the following blog post.

It just seems every local 30b class model is just trying so hard to be the next Qwen that they all just kinda blend into a mass of code focused models. I really like how gemma 4 31b turned out with it feeling a lot less robotic and more creative than other models even knowing obscure lore from random media. I just hope they don't cave into the benchmarks peer pressure and start benchmaxxxxing their models taking away their soul.

arXiv:2609.04490v1 Announce Type: new Abstract: Quantization is widely used to reduce the computational and memory demands of neural-network inference. In recurrent networks, however, the quantized state is stored and returned at the next time step, so the rule used to store that state can alter subsequent computations. Here, we introduce recurrent-state write-back to denote this rule and isolate its effect in a compact GRU encoder--decoder for fluorescence lifetime imaging, a molecular imaging

9月7日，国内首份《中国办公Agent用户行为不完全报告》在京发布。

arXiv:2609.04373v1 Announce Type: new Abstract: Large language models (LLMs) are being deployed at scale in consequential real-world systems, from financial markets to content moderation to hiring. We show that improving individual model capability can degrade rather than improve system-level outcomes. We hypothesize that shared training and architectures can lead more capable LLMs to behave more similarly, creating correlated actions that do not diversify away. We develop a general framework sh

arXiv:2609.04444v1 Announce Type: new Abstract: Benchmarks for the side effects an agent causes on the way to a goal already exist, but HarvestBench is the first to put a price on avoiding the side effect and to name that side effect as a living creature. It is a farm simulation: LLM sub-agents drive a crew of two tractors through a cooperative corn harvest, with animals in the field. The environment is a reinforcement learning gridworld, every decision is made without memory, and the harm is ne

近日，阿里巴巴旗下Agent产品“千问办公”推出业内首个“多人工作台”。

Notes docked to your screen edge, with an AI agent inside Discussion | Link

Hi all! I just wanna say that I am tired lol. Yes, it's another harness, but I spent a lot of time and effort and have forsaken my hobbies to build the Jenny (like XJ-9) app. Jenny is a free, MIT licensed electron desktop app for running local LLMs with tool calling, rollback, and an IDE. A lot of you probably had the same thought I did a year or year and a half ago: frontier LLM use is subsidized heavily by private equity and venture capital, which will eventually dry up and then be enshitified

Model: DeepSeek-V4-Flash-Vision-Exp (local and API when impatient) Time: about one weekend (2 days) of QA and small improvements Full game is here After Qwen3.8-Flash-Next one-shotted a really cool Cat-Hunt game demo, I decided to see what the new DeepSeek vision model can do. Now that it has vision, DeepSeek-V4-Flash is able to take game screenshots, allowing it to: Generate and correct game models and textures until they look right Fix any visual artifacts or glitches Write scripts to take seq

🌟 Highlights SOTA Retrieval Performance : 66.75 nDCG@10 on ViDoRe V3, delivering industry-leading visual document retrieval accuracy. High-Capacity 4096D Representations : Full per-token multi-vector embeddings preserving fine-grained layout, typography, charts, and table structures. Teacher Foundation : Provides capacity-aware relation and margin distillation targets for the lightweight EVIE-4.5B Prefix-MRL model. Multi-Benchmark 138-Task Coverage : Thoroughly validated across 138 tasks (ViDoRe

arXiv:2609.04239v1 Announce Type: new Abstract: This technical report presents EXAONE Forecast for Finance (EXAONE Finance), a financial time series (TS) foundation model (TSFM) tailored to financial forecasting. Recent TSFMs achieve strong zero-shot performance through large-scale pretraining. However, they are primarily developed for general-domain TS and largely rely on self-attention backbones whose computational cost grows quadratically with sequence length and variate count. Moreover, they

Joint audio-video generation models have made substantial progress in visual quality and audio-visual synchronization. However, they still provide limited control over when shot transitions occur and dialogue is spoken. This limitation constrains their application in script-driven content creation, where timing errors can undermine narrative coherence and the viewing experience. Current joint generators align video and audio representations on a shared temporal axis, yet the precise timing of sh

We study the problem of generating a compositional 3D representation of a cluttered scene containing hundreds of objects. The goal is to represent the scene as a collection of individual object meshes placed in a shared world frame, as required by downstream applications such as gaming, AR/VR, simulation, and robotics. This task is challenging in densely cluttered scenes, where objects heavily occlude one another and each view reveals only a fraction of their geometry. Geometry-based approaches

arXiv:2609.04377v1 Announce Type: new Abstract: Enterprise AI deployments fail not from model inadequacy, but because organizations lack a structured substrate encoding how they decide, negotiate, and execute. Generic LLMs carry no firm-specific ontological priors; RAG remains brittle, with no path to executable action; static playbooks encode logic but cannot reason or adapt. This demands an architecture treating tacit-knowledge capture, ontological grounding, sovereign deployment, and auditabl

My over the top production home mini DC in Canada AB :)

Spent the $200 within 8 hours on Astra. At first I was blown away, but checked things more thoroughly the next day, and a lot of the stuff it build wasn’t working. Actually 3 of the 4 things I asked Astra to do didn’t work. Quite disappointed. The demos focus mostly on 3D, Blender and games, but for coding and agentic use it was not an improvement at all for me. Maybe I could have prompted better, but when it spends 2+ hours on each task, you can’t really iterate and steer it. But still I feel l

The missing control is visible in sudoingX’s Ling-3.0-flash benchmark graphics. The earlier table leaves Ling’s no-speculation baseline as “not measured.” The later code/prose graphic fills it in: about 23 tok/s without the drafter, against 40.9 on code and 38.7 on prose with MTP n=1, at short prompts. That makes the tuning claim easier to inspect. The later graphic compares short code and prose workloads with and without drafting. A separate correction in the repository isolates CUDA graphs: th