Open-Weight LLM Showdown Week 7: Mistral Small 4 Impresses but Stays Out of Reach
Mistral Small 4's 119B MoE unifies reasoning, vision, and coding—but needs datacenter hardware. Qwen 3.5 35B-A3B remains the consumer GPU king at 112 t/s.
Tag
#llm
Open-Weight LLM Showdown Week 6: MiMo-V2-Flash Brings 309B Parameters to Consumer GPUs
MiMo-V2-Flash runs 309B parameters on RTX 4090s. GLM-5 sets benchmarks but needs datacenters. Llama 4 Scout stays out of reach.
Open-Weight LLM Showdown Week 5: Qwen 3.5 Dominates, Nemotron 3 Super Redefines Efficiency
Qwen 3.5's MoE models hit S-tier benchmarks, NVIDIA's Nemotron 3 Super delivers 5x throughput gains, and GLM-4.7-Flash brings frontier coding to consumer GPUs. The open-weight race just accelerated.
Your Customer Service Chatbot Is a Security Hole
IEEE S&P research finds 10,000+ websites running vulnerable AI chatbot plugins. Attackers can forge conversations, hijack tools, and extract system prompts.
Open-Weight LLM Showdown: GTC Pivots to Inference, DeepSeek V4 Still MIA
Jensen Huang bets on inference chips, Ollama adds multimodal support, and DeepSeek V4 remains the most anticipated release that hasn't happened yet.
Open-Weight LLM Showdown: Mistral Small 4 Arrives, DeepSeek V4 Finally Lands
Mistral drops a 119B MoE model under Apache 2.0, DeepSeek V4 emerges from stealth, and dual RTX 5090 setups are matching H100 on 70B inference. This week changed the game.
Yann LeCun's $1 Billion Bet: World Models Will Beat LLMs
After 12 years at Meta, the Turing Award winner raised the largest European seed round ever to prove the AI industry got it wrong.
Best Local Coding Models by VRAM Tier (August 2026)
Which open-weight coding model to run locally? HumanEval and SWE-bench picks from 8GB to 32GB GPUs. Qwen2.5-Coder, Qwen3.6, Devstral, KAT-Coder.
Best Local Chat Models by VRAM Tier (August 2026)
Head-to-head comparison of local chat and assistant models from 8GB to 32GB VRAM. Current picks: Qwen3.5, Gemma 4, GPT-OSS, Qwen3.6, and GLM-4.7-Flash.
Open-Weight LLM Showdown: RTX 5090 Finally Delivers, But You Can't Buy One
Two weeks after our last roundup, the 5090 benchmarks are in and Qwen 3.5 Small models are running on phones. Here's the real performance picture.
Grok 4.20 Turns AI Into a Debate Team: Four Agents Argue Before Answering
xAI's new multi-agent architecture pits four specialized AI agents against each other in real-time debate, claiming 65% fewer hallucinations
Your Anonymous Internet Account Isn't Anonymous Anymore
New research shows AI can link your Reddit burner account to your real identity for under $4. The era of casual online pseudonymity may be ending.
Your Burner Account Won't Save You: AI Can Unmask Anonymous Users for $4
New research from ETH Zurich and Anthropic shows AI can identify pseudonymous online accounts with 67% accuracy at just $1-4 per person - and there's no easy fix.
MiniMax M2.5: The Open-Source Model That Rivals Claude at 1/20th the Cost
Chinese AI startup MiniMax has released M2.5, an open-weights model matching Claude Opus performance for coding and agentic tasks while costing 95% less to run
Open-Weight LLM Showdown: What Runs on Your GPU (October 2026)
GLM-5, Qwen 3.5, DeepSeek V3.2, and MiniMax M2.5 are rewriting the rules. Here's what they actually deliver on consumer hardware.
PsychAdapter: The AI That Can Fake Your Personality With 98% Accuracy
Researchers have created an AI system that can generate text mimicking specific personality traits and mental health conditions. The implications for manipulation and misinformation are troubling.
The Invisible Censor: How China's AI Chatbots Learn to Forget
Stanford and Princeton found Chinese AI models refusing sensitive questions up to 60% of the time, against under 3% for Western models
MIT's AI Learns Yeast DNA Language for Drug Manufacturing
An LLM trained on yeast genetics outperforms commercial tools at codon optimization. Five of six test cases beat existing solutions.
Google+UVA: Longer Reasoning Predicts Failure, Not Success
Google and UVA research shows longer AI reasoning traces correlate with wrong answers. The fix: measure how deeply the model thinks, not how much it writes.
AI Chatbots Beat Human Teams on Medical Data, UCSF Finds
Half the tested AI tools produced prediction models that matched or beat human researchers. A student and high schooler built working code in minutes.
Most AI-Generated Passwords Fail Basic Strength Tests
Kaspersky finds DeepSeek, Llama, and ChatGPT all produce password outputs that fail standard strength tests. Prediction capability makes LLMs bad at randomness.
DeepRare AI Diagnoses Rare Diseases Faster Than Doctors
Shanghai researchers built DeepRare, an AI system using 40+ tools that identifies rare diseases 64% vs 55% for experienced physicians on first attempt.
GLM-5: China's 744B Open Model Trained on Huawei Chips
Zhipu AI releases GLM-5 under MIT license, a frontier model rivaling Claude and GPT-5 while proving China can build top-tier AI without NVIDIA hardware.
AI Models Accept Medical Misinformation 32% of the Time
Mount Sinai researchers tested 20 LLMs with over a million prompts and found they readily accept false medical claims embedded in clinical-looking documents.