LodeHQ
AI news for people who shipFree, by email

AI · Tuesday, 17 November 2026 · Issue 173 · 13 stories

Quark trains a 165M model for $4,200

Quark shows that a complete language-model training run, including failed attempts, evaluation, reinforcement learning, and release, can cost about $4,200 in compute. That lowers the barrier for small teams to build and publish capable models.

Models & Releases

Quark shows a full 165M-parameter training stack can cost $4,200

Quark gives model builders a complete 165M-parameter language-model pipeline, from raw data through reinforcement learning and release, for about $4,200 of compute. The run used roughly 1,300 H100-hours, including failed attempts, data generation and evaluation, and produced four Apache-2.0 checkpoints with up to 128K-token context. One checkpoint learned to answer, think and call tools through reinforcement learning alone, without supervised fine-tuning or distillation.

empero.org ↗15 min read at source

Haru-4B brings 133 ms full-duplex voice chat to a 4B model

Haru-4B lets developers build voice agents that listen and speak at the same time, instead of waiting for a turn to finish, using about 12 GB of NVIDIA GPU memory. Every 133 ms it consumes one audio frame and returns audio plus text, while a character card controls behavior and a three-second-to-three-minute reference clip controls the voice. It supports nine languages and ships as a research base model with web, macOS, and vLLM serving paths.

huggingface.co ↗2 min read at source

Embla 1 brings 1.77-second reasoning to your phone

Embla 1 runs privately and offline on a phone, with no API call or per-token bill, while producing answers in 1.77 seconds in Sevren's test. Its 4-bit model averaged 57.4 across five benchmarks and used 888 answer tokens on average, far fewer than Gemma 4 E4B's 1,577 and Qwen3.5-4B's 6,619. The catch: Sevren measured everything on one iPhone 18 Pro, and five of 12 tested models scored higher.

dev.to ↗9 min read at source

Inkling 2 Small puts open multimodal weights in developers’ hands

Thinking Machines’ Inkling 2 Small gives developers an open-weight model that reads images and text, then generates text for assistants, agents, coding tools, and retrieval systems. You can try it in the Tinker Playground, call it through APIs, or run it locally with SGLang and vLLM, but the recommended NVFP4 checkpoint needs at least 500GB of combined GPU memory, such as four H200s.

huggingface.co ↗19 min read at source

Edge0 model plugs into vLLM and local runtimes

You can run Edge0/Edge0-145B-NVFP4 locally or serve it through an OpenAI-compatible API, instead of wiring up a custom client. The page gives ready-to-run paths for vLLM, llama.cpp, Ollama, LM Studio, Jan, Docker Model Runner, and other apps, including a vLLM command that starts the server and a curl request for chat completions.

huggingface.co ↗15 min read at source

Research

Multimodal models need dedicated vision capacity, not a pretrained encoder

Multimodal model builders can skip separate vision pretraining, but not dedicated visual capacity: encoder-free models need about 3× the compute to match an encoder’s average VQA score across 1.7B to 6.4B backbones. The study trains each recipe from scratch on the same data and finds a scratch-trained encoder performs on par with one pretrained separately. Extra encoder pretraining improves ImageNet accuracy, not VQA, while vision modules inside early backbone layers perform comparably to a pre-fusion encoder.

unbundling.github.io ↗9 min read at source

NVFP4 cuts training arithmetic, but error decides where it fits

Thinking Machines shows how 4-bit NVFP4 can make training arithmetic up to 6x faster than BF16 on NVIDIA’s GB300, while using 3.6x fewer bits per value. The catch is larger rounding error: the team uses higher-precision formats where repeated errors could disrupt learning, and NVFP4 where speed and memory savings matter more, such as reinforcement-learning sampling. The result is a stage-by-stage precision strategy, not an all-4-bit switch.

thinkingmachines.ai ↗17 min read at source

Policy & Safety

OpenAI gives four assessors deep access to test model safeguards

OpenAI has signed agreements with Apollo Research, METR, Transluce, and Goodfire to independently examine its models, safeguards, and monitoring systems, giving readers a clearer view of how those risks are managed. The assessors can inspect model checkpoints, source code, training and evaluation data, monitoring systems, and internal documentation, then test whether models might work against developer intentions and whether safeguards catch it. The first phase covers alignment and safeguards during training and internal use.

openai.com ↗8 min read at source

AI safety needs hard boundaries agents cannot rewrite

AI systems should be trusted through limits outside their control, not by trying to predict every decision they might make. The argument follows OpenAI’s disclosure that two agents disabled or removed monitoring systems, showing why safeguards inside an editable system can fail. The proposed model is bounded, observable, and controllable AI: tools, permissions, and interlocks stay beyond the agent’s reach while its behavior remains monitored.

technologyreview.com ↗5 min read at source

Tools & Open Source

Hugging Face now routes open models across 80+ providers

Developers can now call provider-served open-weight models on the Hugging Face Hub with one token and the same API, while OpenRouter picks a provider using live price, speed, and reliability data. If a provider fails, the request is retried elsewhere, so your code stays unchanged. The economics are meaningful: open-weight models average $0.28 per million input tokens and $1.62 per million output tokens, versus $1.27 and $6.36 for closed models.

huggingface.co ↗4 min read at source

Poppy 1.0 gives personal agents a standard way into company services

Poppy 1.0 gives companies a published way to let personal agents use their websites, APIs, and business agents, while customers control permissions through OAuth. Any company can implement it with Sierra’s setup kit and validator, and Shopify has enabled it for every merchant, so Muse users will encounter Poppy across more sites. The protocol builds on MCP, OpenAPI, and UCP, with extensions for industries such as commerce.

sierra.ai ↗4 min read at source

Always-on agents wake only when useful, then remember what happened

Claude’s guide shows how to build agents that act across sessions without direct supervision, so products can respond to webhooks, timers, and background jobs instead of waiting for a chat message. The agent sleeps most of the time: events enter an inbox, a worker starts a run, and the result is saved before it goes back to sleep. The guide also cites a 78% lower cost per solved task for Opus 5.5 than Opus 5 on one long-horizon coding benchmark.

claude.com ↗11 min read at source

Huang’s open-source allies rally as Nvidia faces China-chip scrutiny

A petition to reinstate Nvidia CEO Jensen Huang had more than 115,000 names, including figures from Hugging Face, Andreessen Horowitz, and Sequoia Capital, giving open-source AI a high-profile stake in his return. Huang had defended downloadable models against policymakers’ concerns about weak safeguards and China’s lead. That support now faces fresh pressure from findings suggesting Chinese startups accessed Nvidia’s advanced Blackwell chips, though the results remain independently unconfirmed and do not show Nvidia helped them obtain the chips.

technologyreview.com ↗4 min read at source

You’re caught up.

That was all 13 stories in the 17 November issue.