LodeHQ
AI news for people who shipFree, by email

AI · Sunday, 15 November 2026 · Issue 171 · 14 stories

A 145B agent now runs on one consumer GPU

Edge0-145B puts a 145-billion-parameter agent on one RTX 5090 D, giving developers near-datacenter scale without datacenter hardware. Its sparse design reports 98.3 tokens per second while using 23.25 GiB of VRAM.

Models & Releases

Edge0 fits a 145B agent on one consumer GPU

Edge0-145B gives developers a 145B-parameter agent that can run on one RTX 5090 D, instead of requiring a datacenter setup. Its NVFP4 build uses 23.25 GiB of VRAM plus 38.27 GiB of host RAM, and the project reports 98.3 tokens per second for agent decoding. The model is a sparse mixture of experts: 10 of 512 experts are active for each token, keeping the full model large while limiting work per step.

huggingface.co ↗5 min read at source

LongCat-2.5 targets long-running coding and visual agents

Meituan’s LongCat-2.5 is a 1.6-trillion-parameter multimodal model for agents that can work across codebases, visual tasks, creative design, and 3D work. It turns native-resolution images into discrete tokens and trains on real engineering tasks plus reproduced agent failures, rather than treating vision as a separate add-on. The model reports 88.4 on Terminal Bench v2.1 and 86% higher per-device throughput than its DDP training baseline on long-sequence SFT tasks.

huggingface.co ↗10 min read at source

Granite’s LLMON previews sharply cut prompt injection success

IBM’s Granite 4.1 3B and 8B LLMON previews give developers a way to separate instructions from data, making indirect prompt injections nearly fail on some benchmarks instead of hijacking the model with high probability. LLMON wraps meaningful tags in reserved tokens, adds explicit nesting and unique IDs, then binds each instruction to the data it should use. IBM also released a toolkit, examples, and JSON converters for testing the format.

huggingface.co ↗9 min read at source

Decider v1.3 turns model judgments into usable probabilities

Perplexity’s Decider v1.3 and smaller Decider Flash v1.3 let developers classify, route, rank, and evaluate text and vision inputs without generating free-form answers. You give the model an input plus questions with possible answers, and it returns probability distributions in one pass for yes/no, multiple-choice, or ordered score decisions. Both models are available on Hugging Face under Apache license, through OpenRouter, and via Perplexity’s Decisions API.

perplexity.ai ↗19 min read at source

LiquidCodec makes on-device speech generation faster and lighter

LiquidCodec gives Liquid AI’s newest on-device audio models a codec built for fast, local conversation, not just high-fidelity playback. It turns speech into language-model tokens, then decodes predicted tokens back into audio through an encode, quantize, decode pipeline. At roughly 74 million parameters, it matches similarly sized codecs on reconstruction quality and compression while outperforming them in speech generation and token-to-waveform decoding speed.

liquid.ai ↗14 min read at source

Decision models answer fixed-choice tasks without generating text

Hugging Face’s decision-model support lets applications classify inputs into predefined choices and receive probabilities directly, instead of asking an LLM to generate structured text first. A Transformers pipeline takes a state, such as a duplicate-charge refund request, and returns the selected department plus confidence and probabilities. In the post’s example, it picks billing with 0.9974000453948975 confidence and assigns billing a 0.9982666969299316 probability.

huggingface.co ↗15 min read at source

Research

XYBench tests whether AI can solve the real problem

XYBench gives researchers 8,115 questions designed around misconceptions, so they can test whether an AI assistant fixes the user’s underlying problem instead of blindly answering the request. In one example, a user asks how to make frosting stick when the cupcakes are too warm; the right response addresses the heat first. Five evaluated models often stayed focused on the mistaken approach, unlike human reference answers.

huggingface.co ↗6 min read at source

Token pooling keeps 97.8% of retrieval quality at one-quarter storage

A pooling-aware training method lets multi-vector search models cut document vectors 4x while retaining 97.8% of their unpooled NanoBEIR nDCG@10, up from 94.6% with normal training. It teaches the encoder what information will survive compression, then merges similar token vectors with hierarchical clustering. In the post’s example, pooling 200 vectors to 50 cuts quantized storage for one million documents from 12.8 GB to 3.2 GB.

huggingface.co ↗12 min read at source

HOPE prunes half of MoE experts while preserving expert teamwork

Amazon’s HOPE method lets MoE model builders remove 50% of their experts with less performance loss, improving average rank and gains of up to 6.1% on agentic coding tasks versus the next-best pruning method. Instead of scoring experts one at a time, it scores pairwise interactions, so experts that work together are less likely to be separated. The approach needs no training, avoids dynamic loading, and works with quantization.

huggingface.co ↗9 min read at source

Claude’s oldest text candidates are Sumerian accounting records

Claude’s training corpus may reach back to Sumerian accounting entries from 3400 to 3000 BC, giving researchers a strange way to probe what “text” means in a modern model. The records track debt and commodity logistics, rather than literature; the best-known example is the Kushim Tablet, which records an exchange of 29,086 measures of barley over 37 months. Archaeological dating is too broad to identify the single oldest tablet.

lesswrong.com ↗7 min read at source

Policy & Safety

Automated alignment splits into three jobs with different risks

Automating alignment is not one strategy: it breaks into empirical research, conceptual research, and direct work on making AI systems more aligned. The first is the clearest near-term fit because agents can design and run alignment experiments, but it shares heavily with automating AI capabilities research, so accelerating it may not be the priority. The harder gap is conceptual work, including choosing metrics, because progress lacks a clear, verifiable signal.

alignmentforum.org ↗13 min read at source

AI protesters want Trump and Xi to halt the arms race

Stop The AI Race is asking Donald Trump and Xi Jinping to negotiate an international treaty that bans development of more powerful AI systems until it can be done safely. The campaign says earlier protests helped push Anthropic and OpenAI to support an internationally coordinated slowdown, but produced only voluntary White House commitments. Its next target is the leaders’ Nov. 18 meeting.

aitreatymarch.org ↗1 min read at source

Tools & Open Source

Helion cuts kernel tuning with compile-time performance guesses

Helion can now choose kernel settings at compile time, so developers avoid the long autotuning runs that previously took up to 25 hours for one paged-attention kernel across 72 shapes. It reads the kernel’s intermediate representation and dataflow, classifies it as matmul, reduction, or pointwise, then selects block sizes, warps, and stages. On H100, the approach was 2.13x faster than Helion’s old no-autotune default for linear attention and 2.2x faster across six vLLM reduction kernels.

pytorch.org ↗14 min read at source

DeepSeek turns search, files, and voice into callable API tools

Developers can now add the core DeepSeek App experience to their own products, instead of rebuilding search, file reading, link fetching, safety checks, and voice features themselves. V4.1-Flash handles tool calls and context assembly server-side, while standalone Search, Fetch, OCR, ASR, and TTS APIs cover custom workflows. Public-facing apps must enable safety checks, and tool fees are charged alongside standard token costs.

api-docs.deepseek.com ↗1 min read at source

You’re caught up.

That was all 14 stories in the 15 November issue.