← Daily LLM news

Daily LLM news, Sunday, 15 November 2026

LLM Daily Briefing: Sunday 15 November 2026

Hugging Face logo and brand image
Image: Hugging Face (huggingface.co)

Good morning, here's today's briefing. The lead: Edge0 releases open weights for a 145 billion parameter model it says runs on one consumer graphics card. Also today: Perplexity's Decider models learn to pull exact text out of documents, llama.cpp reaches version 1.0, DeepSeek opens its app's search, file and voice tools to developers, and Tencent launches a real-time voice model that can call tools and sing.

LEAD: Edge0 releases a 145 billion parameter model built to run on a single consumer GPU

Background: Large open models usually need data-centre hardware. Mixture-of-experts designs help, because only a small slice of the model works on each word, but the whole model still has to sit somewhere in memory.

What happened: Edge0 published Edge0-145B and a smaller Edge0-120B on Hugging Face, along with its inference software. The model card describes a mixture-of-experts model built on Qwen3.8-Flash-Next, with 145 billion parameters, about 10 billion active per token, a 256K context window and native 4-bit training. Edge0's engine keeps the most-used experts on the graphics card and streams the rest from system memory. On an RTX 5090 D with 24 GB, the NVFP4 build used 23.25 GiB of video memory plus 38.27 GiB of system RAM, and Edge0 reports 98.3 tokens per second when decoding agent tasks and a Terminal-Bench 2.1 score of 86.5. It also lists 57 tokens per second on an Apple M2 Ultra. AIVIO News reports Edge0's internal score of 64.6 on the DeepSWE 1.1 coding test, ahead of Qwen3.8-Flash-Next at 58.7. All benchmark figures are Edge0's own. The licence allows personal, research and internal use, but selling the model, running a commercial model service or offering a commercial coding or office assistant with it needs a separate Edge0 licence. Edge0 model card, AIVIO News

Why it matters: If the speed holds up outside Edge0's tests, a model of this size on a desktop card puts serious local coding agents within reach of individuals. The licence, though, keeps it out of most commercial products for now.

Perplexity's Decider v1.3 can now pull exact text spans out of documents

Background: Perplexity's Decider models do not write answers. They return probabilities for yes or no, multiple-choice and scoring questions, which makes them cheap tools for routing, classifying and ranking.

What happened: Perplexity released pplx-decider-v1.3 and pplx-decider-flash-v1.3, with weights under Apache 2.0 on Hugging Face, MarkTechPost reports. The main addition, Locate, returns exact text spans with character positions, for jobs such as picking out names or personal data. The larger model is about 27 billion parameters on a Qwen3.8-27B base, and Flash about 4 billion. In Perplexity's own results, v1.3 averages 91.6% on 63 query-classification tasks and flips its answer on only 0.27% of reordered or relabelled questions. Locate is its weakest area, averaging 69.0. API pricing is $0.02 per million input tokens for v1.3 and $0.01 for Flash, with output free. MarkTechPost

Why it matters: Not every step in an agent needs a frontier model. Small, open decision models like this can handle sorting and checking at a fraction of the price.

llama.cpp reaches version 1.0 after three and a half years

Background: llama.cpp and its tensor library ggml are the engine underneath a large share of local AI apps, running models on everything from laptops to servers.

What happened: The team, now at Hugging Face, released llama.cpp and ggml version 1.0.0 on 13 November. From now on both follow semantic versioning: a major version bump means a breaking change to the public C interface, a minor one means compatible additions such as new model support, and a patch means bug fixes. The team says llama.cpp supports more than 150 architecture variants, and that hardware support from Nvidia, AMD, Qualcomm, Arm, IBM and Intel is helping it run tests across platforms. Hugging Face, GitHub release

Why it matters: A stable interface means apps built on llama.cpp should break less often when it updates, which matters as more products depend on running models locally.

DeepSeek opens its app's search, file and voice tools to developers

Background: Chat apps bundle features such as web search, file reading and voice, while developers using the raw API have usually had to build those parts themselves.

What happened: DeepSeek announced on 13 November that the capabilities behind its app are now in its API. Its V4.1-Flash model supports web search with citations, search during multi-step reasoning, file input for PDFs, Office files and images, automatic reading of links, and safety checks, with tool calls handled on DeepSeek's servers. Search, fetch, OCR, speech recognition and text to speech are also offered as standalone APIs. DeepSeek says public-facing apps must turn on safety checks, and the extra tools carry fees on top of token charges. It also released an open-source demo front end. DeepSeek

Why it matters: This lowers the cost of building a full chat assistant on DeepSeek, and puts it in more direct competition with the built-in tool suites of OpenAI and Google.

Tencent launches Hy Realtime, a voice model that can call tools and sing

Background: Real-time voice models are becoming the front door for assistants and customer service lines, where speed and natural speech matter as much as raw intelligence.

What happened: Tencent's Hunyuan team released Hy Realtime on 13 November, according to a report carried by Tencent News. Tencent says it performs well on speech understanding, following instructions and natural expression, and adds agentic tool calling and singing. It can decide from the conversation whether to run a web search. Hy Realtime is already live in voice calls in Tencent's Yuanbao app and in Tencent customer service, and Tencent plans to offer it as an API on Tencent Cloud soon. Tencent did not publish benchmark figures in the report. Tencent News

Why it matters: Voice assistants that can look things up mid-call are moving into everyday Chinese consumer products, and a cloud API would put Tencent up against OpenAI and Google's realtime voice offerings.

← Friday, 13 November 2026

Get the weekly sift

The best new AI and agent releases, with one-line verdicts. Occasional picks, no spam.

Privacy policy: your email goes to our email provider, Resend. Unsubscribe any time.

News is editorial and never paid for. Sponsored content on ToolSift is always labelled.