Tech news & AI articles - LinkWord
AI and tech news, product notes, and practical tutorials—curated long reads updated regularly.
GLM-5.3: Zhipu AI's Coding Model Leaps Forward with Unexpected Cyber Capabilities | 2026-08-20
Zhipu AI released GLM-5.3, achieving a 50% coding improvement and open-source SOTA on cyber vulnerability discovery through post-training RL scaling alone.
8/20/2026
10 views
Qwen3.8-27B Open Weights Released: 27B Model Beats Claude Opus 4.6 Max on Key Coding Benchmarks | 2026-08-15
Alibaba open-sourced Qwen3.8-27B on Aug 14: a 27B dense multimodal model with native image/video understanding and 262K context (up to 1M). It beats Claude Opus 4.6 Max on several coding benchmarks (SWE-bench Pro 61.7, DeepSWE 42.2) and runs on consumer hardware. DeepSeek also shipped V4-Pro GA plus peak/off-peak pricing.
8/15/2026
14 views
Gemini 3.7 Flash, DeepSeek Harness, GPT-5.6 Sol Ultrafast, Mistral OCR 4.1: AI Roundup | 2026-08-14
Four AI releases landed on August 13: Google's Gemini 3.7 Flash at half the price, DeepSeek's open-source Harness agent framework (MIT, everything is a plugin), OpenAI and Cerebras' GPT-5.6 Sol Ultrafast tier (up to 750 tokens/s), and Mistral OCR 4.1.
8/14/2026
7 views
Grok 4.6, DeepSeek V4 Pro 0813, and Open-Weight Qwen3.8-2.4T: Three Major AI Releases | 2026-08-13
Three major releases in 24 hours: xAI's Grok 4.6 matches GPT-5.6 Sol on the AA Intelligence Index (61), DeepSeek ships the official V4 Pro 0813 with much stronger agent capabilities, and Alibaba Qwen open-sources the 2.4T-parameter Qwen3.8-2.4T-A95B — the largest open-weight model ever.
8/13/2026
21 views
NVIDIA Launches Nemotron 3.5 Lightning: Open 30B Agentic Model with 4x Faster Output | 2026-08-12
On August 11, NVIDIA released Nemotron 3.5 Lightning, an open 30B (3B active) MoE agentic model with a Mamba-2 + MoE + Attention hybrid architecture and up to 1M token context, plus the open-source NeMo Switchyard routing library. NVIDIA claims up to 4x faster output and ~30% faster agentic task completion. Available on Hugging Face, ModelScope, OpenRouter, and as a NIM microservice; runs locally on RTX PCs and DGX Spark under the commercial-friendly OpenMDW-1.1 license.
8/12/2026
12 views
Meta Muse Glimmer: 30B Open-Weight Local Agent Model, Under 20GB Quantized | 2026-08-11
Meta Superintelligence Labs open-sources Muse Glimmer, a 30B Apache 2.0 agentic model for on-device use. Distilled from the flagship Muse Spark and quantized to under 20GB, it runs on a single consumer GPU. Meta claims it beats Gemma4-31B and Qwen3.6-27B on agentic benchmarks, with DFlash speculative decoding delivering 3.1x faster generation on RTX 5090.
8/11/2026
14 views
DeepSeek V4 Flash 0731 ARC-AGI Results: 89% on ARC-AGI-1, 61.4% on ARC-AGI-2 | 2026-08-09
ARC Prize published verified ARC-AGI results for DeepSeek V4 Flash 0731: 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2 at max effort, costing $0.02-$0.04 per task — comparable to GPT-5.6 Luna at roughly 2x lower cost. Official agent benchmarks far exceed V4-Pro-Preview; open weights run on a single MI300X. DeepSeek also announced a significant API price increase is coming.
8/9/2026
13 views
US DOE Launches Genesis Open Models Initiative with Trillion-Parameter Genesis-Science-1 | 2026-08-08
The U.S. Department of Energy launched the Genesis Open Models Initiative with Arcee AI, unveiling Genesis-Science-1, a trillion-parameter-class open-weight model for scientific research across materials, energy, fusion and biology. Weights and a technical report are slated for release later this year.
8/8/2026
12 views
Qwen3.8 Max Tops Agentic Index, GPT-5.6 Luna Goes Free: AI News Roundup for August 6 | 2026-08-07
On August 6, Alibaba's Qwen3.8 Max was ranked first overall on Artificial Analysis' Agentic Index, with 2.4T parameters and open weights coming for the first time at Max scale. The same day, OpenAI updated GPT-5.6 Sol and made Luna the default model for free users with unlimited text chats.
8/7/2026
15 views
DeepSeek V4 Flash Runs on a Single AMD MI300X: 304B Parameters, No Quantization | 2026-08-05
Developer ryanzhou open-sourced configs and patches to run DeepSeek-V4-Flash-0731 (304B params) on a single AMD MI300X, with no quantization or weight offload, hitting 168.6 tok/s single-stream decode. The HN post scored 365 points the same day.
8/5/2026
8 views
Mistral Releases Shieldstral: 3B Open-Weight Multimodal Content Safety Classifier | 2026-08-05
On August 4, Mistral released Shieldstral, a 3B open-weights multimodal safety classifier under Apache 2.0. It frames moderation as policy-adaptive question answering: policies are supplied as plain-language queries at inference time, returning calibrated safety scores without retraining, unifying text and image safety evaluation.
8/5/2026
13 views
OpenAI's Astra Solves 10 Open Problems in Math and Theoretical CS for ~$2,000 | 2026-08-03
On August 1, OpenAI announced that an internal version of Astra, its next major model family, solved 10 open problems in mathematics, quantum complexity, and theoretical computer science that had seen no progress on their main results for at least a decade. The proofs span sphere packing, non-sofic groups, Connes' rigidity conjecture, and quantum parallel repetition, all formalized in Lean. At Sol API rates, the total token cost to find these solutions was roughly $2,000.
8/3/2026
13 views