Qwen3.8 로컬 실행, 에이전트 하네스 확장, 추론 검증 신호와 문서 추출 한계

Uncensored Qwen3.8 27B specifically for red teaming & security research run on your computer locally, - 262K context - Refusal rate Near-zero : collapsed across AdvBench, HarmBench, StrongREJECT, etc. - Capability: almost fully retained - Ready for real red teaming - MMLU/ GSM8K/ CMMLU barely moved - Vision, tools, reasoning, and 262K context all preserved They didn’t just jailbreak it. They surgically removed the refusal direction from Qwen3.8 27B and re-quantized it to official FP8.

Wtf, Qwen3.8 27B wrote the entire a fully playable 3D FPS game locally in the browser. Setting: elegant Chinese lotus courtyard Targets: floating porcelain drones Ornate weapon. Smooth controls: mouse lock, sprint, reload, hold Z for scope. Everything you see was generated by Qwen3.8 27B (code) https:// x.com/MaziyarPanahi/ status/2088616278470570400/video/1 …

Let Sol manage an efficient fleet of Luna agents for you. These models know each other well and collaborate to achieve the result in an incredibly fast and efficient way.

114,000 GitHub stars in just 2 days. DeepSeek Harness just became the fastest-rising open-source project in GitHub history. faster then OpenClaw/Linux

We’ve written an FAQ to answer some of the questions we've received about watermarking. In summary: • We’re implementing watermarking to comply with the EU AI Act. Other major model developers have signed the same Code of Practice and will also be implementing watermarking;

How will the watermarking actually work? Can it be hidden with editing? And how does this affect code?

8 ways to improve LLM reasoning at inference time: (used by Google, OpenAI, and Anthropic) When accuracy on a task falls short, the cheapest place to intervene is inference. There's no retaining, so the techniques either go in the prompt or the orchestration around the call. Inference compute scales in two ways: - You sample the same prompt several times and select among the results - or you extend one trajectory further before the model answers. Parallel samples batch and run concurrently, so more samples cost tokens and GPU time while it barely affects latency. Sequential tokens each condition on the one before, so more thinking costs latency directly. Tree methods compose both, and inherit both failure modes. 1) Chain of thought The model is prompted to think step by step, so intermediate results land in the context where later tokens can condition on them. 2) Majority voting Sample several traces and return whichever final answer appears most often. Agreement replaces a reward model. It reduces variance and does nothing about bias. When the model misreads a problem the same way every time, the traces agree, and the vote returns that answer with higher confidence. 3) Best-of-N Score N complete answers with a reward model, keep the top one. True reward rises, peaks, then declines as optimization pressure on the proxy increases. Past that peak, the search starts finding answers the reward model likes instead of the answers that are actually correct. 4) Extended thinking The model spends a token budget on internal reasoning first. Longer is not uniformly better. Anthropic built tasks where accuracy falls as traces lengthen, including counting problems where irrelevant numbers pull the model off track. 5) Self-refinement The model critiques and rewrites its own answer. On GSM8K, GPT-3.5 fixed 7.6% of its wrong answers and broke 8.8% of its correct ones. This is net negative, and it works when the critique comes from a compiler or a test suite instead. 6) Tree of thought Propose several next steps, score them, keep the best few, backtrack from dead ends. 7) Beam search Keep K partial solutions and score every step with a process reward model. The technique works better than best-of-N at small budgets and worse at large ones, because the search starts satisfying the scorer instead of solving the problem. 8) MCTS Extend a promising path, roll out to a full answer, push the score back up, repeat. The action space forces a cap on node expansion, and applying that limit leaves the search in local optima. Every technique here generates candidates and picks among them, so the selector sets the ceiling. As a takeaway, build the verification signal first. Scale the search second. To dive deeper, I covered all 8 techniques with the research behind them, the tradeoffs, and pseudo code you can adapt. Read it below.

Jerry Liu

@jerryjliu0

We think a lot about the distribution of all real-world documents, particularly for real-world use cases. We think about the distribution in terms of page length, field length, domain, and task type. Some documents have a single page but can have anywhere from 1-10 fields to 500-1000 fields to be extracted. Other documents are hundreds of pages long and can have 10k-100k fields. On long documents (e.g. a bankruptcy creditor matrix), commercial VLMs lose attention and create large swaths of missing or hallucinated values. The number of fields to extract per page also depends on the extraction task. 1. Needle-in-a-haystack: You find very specific information within a general document corpus. Generally 1-10 fields per page 2. Dense documents: typically extract from filled forms. Quite dense per page, though not super long in terms of total # pages. 3. Long-lists: Documents with both massive #'s of values per page (e.g. 100-1k fields per page) and massive numbers of pages. We'd love for you to explore our overall data taxonomy within ExtractBench, our comprehensive benchmark for real-world document extraction tasks. I love the diagram we show below in Figure 6, that offers one slice of the diverse set of data that we curate! Come check out the ArXiv paper: https:// arxiv.org/pdf/2607.29677 ExtractBench: https:// extractbench.ai

basically, you can generate this level quality of Ai-video locally on RTX 3090 MiniMax H3 runs on your own hardware for text-to-video, image-to-video, and reference-to-video with native stereo audio.

Complete 15s generations on a single 24 GB card with only 31 GB system RAM, -

Andrew Ng’s AI Engineering Skills Map: in 2026 After analyzing 10,000+ job postings and interviewing experts, these are the 4 skills that actually matter right now: 1. Building & deploying AI applications 2. Software engineering fundamentals 3. Using coding agents effectively 4. Shaping the build (product sense) This feels like the clearest roadmap available right now.

"Small-Scale Experiments: Are We There Yet?" Scaling laws show up even in tiny 4M parameter models, but only when their hyperparameters are tuned well enough. This Meta paper found that small models are much more hyperparameter-sensitive. As scale increases, the hyperparameter loss surface becomes lower-dimensional, eventually approaching one effective tuning direction. So the scaling gap may be less about small models failing to predict large ones, and more about under-tuning them. https:// alphaxiv.org/abs/2608.11859 v1 …