2026-08-18

Hi all! I finally just got around to testing out Qwen 3.8 27b. I'm using Unsloth's UD-Q8_K_XL quant as a sit-in replacement to Qwen 3.6 27b, same quant size. Wow -- this thing isn't messing around. I have many baseline test prompts to gauge the 'intelligence' and usability of the model, but a go-to one is asking it to do a 1:1 recreation of classic arcade games (like Galaga, Donkey Kong, Pac-Man, etc). I do this to see what little details it gets correct. I've tested this process on pretty much

Support for the new ling 3.0 models has been merged into llama.cpp: Ling tiny 8b1b - Ling flash 124b5b - Both are reasoning models contrary to prior naming.

llama.cpp is apparently moving to semantic versioning instead of just sequential build numbers (like b10456). The first semantic version tag was created today: Congrats to llama.cpp on version v0.1.0!

Cybersecurity experts who investigate spyware attacks say the number of people who received a recent threat notification from Apple is unusually high.

Nvidia's investment in SoftBank's data center developer will guarantee its chips power an OpenAI data center.

Groq raised $350 million at a $3.5 billion valuation as the former AI chipmaker pivots to a neocloud business and expands its Nvidia-powered data center footprint.

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now DiG-bench shows that Fable displays some creative intuition:…The new frontier for analyzing AI systems is understanding how good they are at inferring the unwritten rules of […]

智谱GLM-5.3首发上线范式PhanRouter，即日开放调用

AI is reshaping cybersecurity for attackers and defenders alike. Learn how OpenAI is strengthening its defenses and what security teams can do now.

The next generation of AI agents is increasingly moving beyond systems that answer isolated questions toward persistent personal assistants that can understand, remember, and continuously learn from users' experiences. Such assistants require long-term memory to accumulate and leverage user-specific experiences over time, yet existing benchmarks remain inadequate for realistic mobile settings, where experiences are heterogeneous, multimodal, evolving, and deeply personal. We introduce MobileMem,

Just wanted to raise some attention to this PR I filed if anyone would like to try it out. This adds an adaptive MTP mode to llama.cpp which employs a fairly simple counting-style state machine to determine the appropriate depth to set the MTP depth to dynamically. The goal here was to make it so people can stop worrying about what the best MTP depth to use is, and just let the server figure it out as it goes. Compared to an MTP of 3, then for very dense hard to predict prose and regular prose t

Hi HN! I'm Bek, founder of Speko, a platform that finds an optimal combination of speech-to-text, LLM, and text-to-speech models, given your constraints, among all our public benchmarked options, and tells you why. Demo: Typical production voice agent is an ensemble of three models: STT, an LLM, and TTS. Each of those layers offers a dozen credible vendors, and each month there are new models on the market. Almost everyone evaluates once, picks a stack of their choice, and never rechecks because

I started building my markdown-svg-renderer tool in May , but I've since added enough features to it that it's worth talking about here again. It's evolved into my ideal tool for sharing Markdown transcripts that include SVG documents. Given my proclivity for drawing pelicans riding bicycles this is a problem that I needed to solve! The tool is very simple. Navigate to markdown-svg-renderer in your browser and paste in some Markdown to see it rendered... or save that Markdown to a CORS-friendly

Agent skills represent a standardized format for packaging procedural knowledge and domain expertise, serving within agent harness systems as an essential mechanism to continually constrain a language model's behavior space for repeatable, high-quality task execution. However, because strong closed-source models entail high inference costs, current popular agent harnesses, such as Codex and OpenClaw, remain prohibitively expensive when deploying these skills to accomplish real-world tasks. The r

Instruction-tuned language models achieve strong performance across a range of generation tasks, but have also recently been shown to exhibit verbalized overconfidence. In question answering, verbalized model overconfidence may be associated with the consistency of the generated supporting rationales. In this paper, we study whether corresponding changes in the lexical diversity of generated answer rationales accompany changes in model confidence induced by instruction tuning. We evaluate three

Hi HN, like many of you talented folk, I’ve been building more, faster than ever. I built Saggar after my terminal stopped feeling like a tool and started feeling like a tab-management problem. Let me know you think! Comments URL: Points: 37 # Comments: 43

arXiv:2608.13612v1 Announce Type: new Abstract: Natural-language interfaces to enterprise data must translate underspecified requests into governed, executable behavior while controlling invalid queries, policy failures, cost, and nondeterminism. SemPlan Benchmark evaluates this architectural design space with a deterministic synthetic bilingual benchmark containing 1,800 cases in English and Brazilian Portuguese; 1,200 cases form the frozen scientific evaluation subset. Four architectures are c

AI最近最出圈的数学突破，都在“找反例”

arXiv:2608.13564v1 Announce Type: new Abstract: Evaluating language-model agents at scale increasingly relies on a second language model as an automatic judge, because the gold signal, an executable environment reward, is expensive, slow, or unavailable at deployment time. Such a judge is a reward-free proxy whose value depends on whether it can be trusted, yet existing judges either hand-write the scoring rubric, as in G-Eval, or fine-tune the judge's weights, and both tend to credit fluent but

Interactive game world models typically autoregress visual observations directly in pixel or latent space, forcing structured properties such as pose, geometry, and occlusion to be implicitly maintained by the same generative sequence. Over long horizons, errors in these latent world properties accumulate, making consistency and controllability fragile. We explicitly model the evolving world state, delegate exact geometric computation to a fixed, zero-parameter renderer, and leave the neural mod

arXiv:2608.13598v1 Announce Type: new Abstract: Agent evaluation relies almost entirely on outcome metrics such as success rate, which capture whether an agent succeeds but not how consistently it behaves. We argue that behavioral consistency across tasks is a distinct and measurable property, and we introduce the Behavioral Consistency Metric (BCM) to quantify it. BCM trains a model to predict task success from behavioral features of agent execution traces, derives a per-trajectory feature-attr

arXiv:2608.13591v1 Announce Type: new Abstract: High-confidence errors in large language models are often treated as evidence of fragile internal inference. We study a different possibility: stable miscalibration, where a confident wrong answer remains locally stable under small perturbations. We combine two diagnostics: a label-aware output-level audit score that ranks domains by confidence variation and overconfident mistakes under a forced-answer baseline, and an internal sensitivity probe th

This Ling 3.0 Tiny 8b param with 1.3b active is the fastest, smartest model I can run on my poor old pc, with 4gb vram. It actually runs lightning fast, like 36 token / sec, as smart as Qwen 3.5 9b / Gemma 12, (Very close), and even faster because of 1.3b active parameters. The Qwen 3.5 9b is running with like 5 token / sec, but this with 36 is finally the speed that i want. Very good open source model, I hope we'll get more of this tiny and really fast models, thank you! :)

Creative agents still lack an effective way to learn from high-quality human films, limiting their ability to produce cinematic-grade videos. A key challenge is the absence of a structured video representation that is both faithful to film content and directly usable for agentic reasoning and manipulation. To address the challenge, we propose the Agentic Video Auto-Encoder (AVA-Encoder), a framework for learning agent-native video representations via agentic auto-encoding. AVA-Encoder transforms

Interactive autoregressive video generation demands both low-latency rollouts and precise online control. Few-step distillation accelerates generation by reducing denoising steps, while online control imposes a causal constraint: frames and blocks should depend on history and controls available during generation. Existing video distribution matching distillation (DMD) pipelines, however, often supervise causal few-step students using bidirectional teachers that score complete clips. The score fo

Sorry for the slop, but I was impressed by this model as I have been testing Qwen3.8-27B Q8_0 locally on my ROG Flow Z13 (Ryzen AI Max+ 395, 128 GB unified memory) and this model was the only one who could made this short simulator (and I have tested a lot of models). Prompt: "Create a beautiful, relaxing flight simulator in a single HTML page." It generated the whole thing through an agent using file/bash tools. Setup: Lemonade Server + llama.cpp ROCm Q8_0 weights + Q8 KV cache Native MTP specu

Qwen 27b Q3_K_M 2x rx 580 8gb (~50$ each in my country, edge cases 60$ per gpu) gives us 16gb vram We used it on an old already existing ddr3 motherboard with 2 gpu slots(you can buy it ror around 200$ with 32 gb of ddr3 ram, a workstation xeon cpu and a workstation motherboard, used) Its not the best option, but it makes running this model possible for many people, its even cheaper than system ram Limitations: very low processing speed(only 14t/s) means an mtp model would be a loss, and high in

Open-source MCP bridge for Blender AI workflows Discussion | Link

Every time I see a post about a newly released model, whether it be a comparison or shitting on it, I have to dig through the endless comments to see what quants they used and what their specs were. Its quite a common occurrence here in this sub to ask someone that's saying a model is underperforming, and when you ask what quantization they are running they say something like "oh im running q0.1bpw from nobodyknowswhothisguyis". Worst offender is with comparison posts. "Comparing the new Qwen3.8

Hi HN. I built 1667 for my own fiction work and now use it each day. This probably has a limited audience. Maybe an audience of one... Why a terminal interface for story writing? I'm a dev. I like to use terminals for a lot of stuff. Most WebUIs feel off to me. That's the only reason. One thing that bothers me about writing in existing tools is that they don't fit the way I write. The mental model of my story is a tree. I try many takes usually continue with just one, but sometimes I want to try

"We have some really ambitious plans to help you work with AI in Chrome to get things done, and I’ll have more to share soon," Jacob Bank, Relay founder and CEO, said.

Higgsfield, founded by former Snap exec Alex Mashrabov, lets users create AI images and videos.

TikTok's parent company signed a memorandum of understanding with The Motion Picture Association.

Apollo did not reach the Moon merely because its engineers could solve difficult equations. It succeeded by turning a distant ambition into a mission architecture of explicit objectives, simulation, verification, and repeated correction. AI now faces a similar transition: frontier models can solve difficult tasks once the problem, tools, and success criteria are specified, yet consequential real-world challenges rarely arrive in an executable or verifiable form. We introduce Apodex Discovery, a

arXiv:2608.13567v1 Announce Type: new Abstract: The human brain exhibits a striking degree of functional specialization, with distinct networks supporting language, formal reasoning, reasoning about other minds, and reasoning about the physical world. Is this modular organization a fundamental principle of how intelligent systems must be built, or an evolutionary accident specific to biological brains? Here, we test whether a similar organization emerges in Large Language Models--another class o

Yes it uses a ton more reasoning tokens than 3.6 did But test in on the same tasks with the other chinese models, glm 5.3, deepseek v4 flash and pro, etc it's really similar, and they are needed The reality is, we're just frustrated because our hardware do not allow most of us to have 1M context (I know that it's not supported yet) with 150 tps decode Furthermore, if you don't mind the quality drop, you can just add a reasoning budget, it will still be better than 3.6

Pose-driven human animation synthesizes a video of a target person from a single reference image and a driving pose stream. Real-time generation is essential for interactive applications such as live streaming, telepresence, and virtual avatars, yet diffusion-based systems require minutes to hours per clip, precluding responsive interaction. We present LiveAnimate, to our knowledge the first animation system to combine real-time streaming with stable long-form generation at billion scale, built

arXiv:2608.13607v1 Announce Type: new Abstract: Frontier LLMs are updated frequently and typically outperform their predecessors in aggregate. But aggregate gains say little about individual samples: an update can still cause sample-level regression, where a response correct under the old model becomes incorrect under the new one. This paper studies how to predict such regressions from signals available at inference time. We compare single-model signals (confidence, logit margin, attention entro

Pretty self explanatory. Could you folks shed some light on why these issues keep happening? I’ve noticed most posts and replies are just people coming to their own conclusions based on whatever published data. I think it would be helpful to get some actual, non-corporate/marketing information on the goings-on by those that actually see what’s happening on the inside. Thanks Can

The web operating layer for AI agents Discussion | Link

Humanoid motion tracking is central to teleoperation and whole-body imitation, yet evaluation often disagrees with what people perceive in videos. Kinematic errors average per-frame pose differences but miss the physical artifacts that matter most, particularly unstable support and incorrect contacts such as foot skating and mistimed touch-downs. Meanwhile, widely used test suites are small and lack the diversity needed to stress contact-rich, long-horizon behaviors. We introduce HumanTracker to

Yes, it sucks to waste time waiting on 16K+ reasoning tokens alone. But here's the thing, this is only a 27B model trying to perform on par with 1T+ parameter models. Something has to be sacrificed, and that sacrifice is the amount of reasoning or trajectory tokens. This isn't new to LLMs whatsoever. Andrej Karpathy himself has said that LLMs need tokens to think. He mentioned this somewhere in his "Let's build GPT" / GPT video series, although I don't remember exactly when, if i am not mistaken

Trying to get a feel for where I stand. If you can list your relevant hardware and model used, that would be awesome. Here's mine: Model: Qwen3.8-27B-heretic-ara, Q5_K_M GGUF T/s by context saturation: I found that t/s output depends on how saturated the context is. The more saturated, the slower the speeds. ~1K | 74.88 tok/s ~16K | 70.15 tok/s ~63K | 56.58 tok/s ~90K | 42.71 tok/s Hardware: 3090 GPU | 64 GBs DDR4 RAM | AMD 7950x CPU Harness: Pi Inference: llama.ccp Edit: When I made this post I