Sakhanda Wire
NVDA $152.56 -2.29% MSFT $505.79 -2.36% GOOGL $279.70 -1.42% META $618.14 -1.32% AMZN $256.03 -0.67%

MarkTechPost

396 stories · all sources

MarkTechPost

FlashMap drafts 16 tokens jointly for faster Qwen3 decoding

Researchers from Polymathic AI, Princeton University, USC ISI, NYU and the Flatiron Institute have released FlashMap, a speculative-decoding drafter for large language models. It uses a continuous flow map to generate a block of 16 tokens in one forward pass, rather than producing them sequentially. The target model checks the proposed block in parallel and retains the longest matching prefix, so the final output remains identical to the target model’s output.

MarkTechPost

Google DeepMind releases Genie 4 real-time world model with native audio

Google DeepMind has released Genie 4 as a research preview for trusted testers. The model creates explorable environments from text and image prompts in real time, generating visuals together with ambient sound, music and character dialogue. Users can navigate with a keyboard and alter weather, camera position, people or objects through plain-language instructions.

MarkTechPost

LEANN cuts a 60-million-chunk vector index to 6 GB

Researchers at UC Berkeley’s Sky Computing Lab have developed LEANN, an open-source vector index designed to reduce the storage required for retrieval-augmented generation and semantic search. Instead of keeping every embedding on disk, it stores a pruned graph and recomputes selected embeddings during a query. In a repository benchmark, a 60-million-chunk Wikipedia index fell from 201 GB to 6 GB. The research paper reports 4 GB for a comparable setup, against 188 GB for HNSW.

MarkTechPost

Liquid AI releases six open-weight LFM3 models

Liquid AI has released LFM3, a family of six open-weight models built from the shared LFM3-1.7B-Base foundation. The 2B-class lineup covers text, vision, audio, multimodal tasks and decision-making. Models include LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B. Each can be fine-tuned on one GPU and deployed on devices, private servers or cloud infrastructure.

MarkTechPost

Mirendil releases Fern 0.1 with a 10M-token context window

Mirendil has released Fern 0.1, its first AI model, designed for AI research and development rather than general conversation. The coding model was pretrained from scratch and built from Mirendil-Base after a short post-training phase. Its context window reaches 10M tokens, which Mirendil estimates can hold about 23 MB of code, or roughly 500,000 lines. The model is available through hosted early access; no open weights or quantized builds have been announced.

MarkTechPost

vLLM Semantic Router Releases 27B d3-ultra Decision Model

The vLLM Semantic Router team has released d3-ultra, a 26.09B-parameter open-weight multimodal decision model licensed under Apache-2.0. It does not generate text. Instead, it reads text, JSON, images or video once per request and returns probabilities for sets of typed questions, including choice, yes/no, scoring, box and tracking tasks.

MarkTechPost

Baidu releases Unlimited OCR 2 for long-document parsing

Baidu has released Unlimited OCR 2, a 3.34B-parameter vision-language OCR model designed to process books, filings and reports spanning hundreds of pages in a single request. The MIT-licensed system converts ordered page images into structured text, including tables, formulas, figures, headings and reading order. It supports a maximum position length of 131,072 tokens and can run on one CUDA GPU in BF16; the tested setup used an NVIDIA A800-80GB, with weights totaling about 6.7 GB.

MarkTechPost

OpenAI releases GPT-6.1 Terra for lower-cost computer-use tasks

OpenAI has released GPT-6.1 Terra, a reasoning model for visual analysis, computer use and writing. It accepts text and images, produces text, and offers a 1.05-million-token context window with up to 128,000 output tokens. The model replaces GPT-5.6 Terra and is available through the API as gpt-6.1-terra. It is also offered in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Its weights are closed and it runs only on OpenAI-hosted services.

MarkTechPost

Nace AI releases Drex for Search for agent document retrieval

Nace AI has released Drex for Search, a retrieval layer designed to help people and AI agents locate evidence across documents. The company says the system can search more than 1M files and can be used by coding agents including Claude Code, Codex, Cursor and OpenCode through a CLI, SDKs, MCP or REST. It can run in Nace’s managed cloud, a dedicated cloud, a customer’s AWS or Azure account, or on-premises and in air-gapped environments.

MarkTechPost

Tsinghua Researchers Open-Source AKV for LLM Agent Cache Management

Researchers from Tsinghua University and CUHK have released AKV, an Apache-2.0-licensed serving system and policy library for engineering KV caches in LLM agents. Built on SGLang and designed for NVIDIA GPUs, it does not include model weights. The system was tested with GPT-OSS-120B, MiniMax-M2.7 and MiroThinker-1.7-Mini across four agent benchmarks.

MarkTechPost

Nanbeige releases 33B open model for agentic coding

Nanbeige LLM Lab has released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license. The dense model has 33B parameters, BF16 weights and a 524,288-token context window. A separate Nanbeige4.3-33B-Base checkpoint is available for custom post-training. The model card’s SGLang and vLLM configurations use four-way tensor parallelism, and the weights require about 66 GB before other memory needs are included.

MarkTechPost

Hugging Face releases open 14.8B DNA model Carbon-X

Hugging Face has released Carbon-X, an open DNA language model with 14.8 billion parameters. Developed by Hugging Face Bio with Zhongguancun Academy and TIGEM / University of Naples Federico II, the decoder-only system can read raw DNA in a chat and respond in English. Its tasks include assessing whether mutations may be benign or pathogenic, identifying regulatory regions, classifying organisms and predicting the next bases in a sequence.