Practical guide for scaling vLLM under OOM and instability
New guide emphasizes workload profiling and tuned configurations over raw hardware for successfully scaling vLLM in production environments.
AINews•2d ago
tutorial
📰 DailyMe
My personalized AI news feed, curated from newsletters and deduplicated automatically.
Powered by OpenHands + Claude Sonnet 4 • Updated every 30 minutes
llama.cpp doom loops with Qwen sampling parameters
Users report llama.cpp experiencing infinite loops with Qwen models at ~20% context even at higher quantization levels, highlighting brittleness of recommended decoding settings across runtimes.
AINews•2d ago
research
Qwen 3.5 Gated DeltaNet avoids KV-cache growth
Gated DeltaNet modules in Qwen 3.5 can avoid KV-cache growth, making it more memory-friendly than Qwen3 under certain operational ratios.
AINews•2d ago
research
NE-Dreamer predicts embeddings instead of reconstructing pixels
George Bredis introduces NE-Dreamer, exploring world models trained to predict next embeddings rather than reconstruct pixels, arguing reconstruction may be wrong objective for control.
AINews•2d ago
research
MMM enables long-context persistent video world models
Gordon Wetzstein introduces Mode Seeking meets Mean Seeking (MMM) as a route to long-context, persistent video world models via unified representation.
AINews•2d ago
research
Beyond Language Modeling: vision-first multimodal pretraining
New paper explores native multimodal models where vision is first-class, treating all modalities Transfusion-style, arguing progress requires vision-native training not language-first adapters.
AINews•2d ago
research
Black Forest Labs previews Self-Flow for multimodal generation
BFL introduces Self-Flow, a self-supervised flow-matching approach for multimodal models (image/video/audio/text) that avoids external pretrained models, claiming 2.8× faster convergence and improved temporal consistency.
AINews•2d ago
research
Cursor now available in JetBrains IDEs via Agent Client Protocol
Cursor announces availability in JetBrains IDEs through Agent Client Protocol, enabling IDE-native access without forcing users to switch tools.
AINews•2d ago
launch
LangChain OSS Skills teach agents to use LangChain/LangGraph
LangChain releases OSS Skills designed to teach coding agents how to effectively use LangChain, LangGraph, and DeepAgents frameworks.
AINews•2d ago
launch
LangSmith Skills + CLI lets agents debug traces and build datasets
LangChain ships LangSmith Skills and CLI enabling coding agents to natively debug traces, build datasets, and run experiments from the terminal.
AINews•2d ago
launch
OpenAI Symphony orchestrates agents via ticket boards
New OpenAI repo Symphony provides orchestration layer that polls project boards and spawns agents per ticket lifecycle stage, shifting UX from prompting to workflow-native automation.
AINews•2d ago
launch
VS Code moves from monthly to weekly shipping
VS Code announces shift from monthly release cycle to weekly shipping of main branch to accelerate feature delivery for agent development.
AINews•2d ago
vendor
VS Code ships agent-focused features for real work
VS Code releases agent hooks, message steering/queueing, integrated agentic browser, and shared memory capabilities designed for production agent workflows.
AINews•2d ago
launch
Windows agent sandbox implementation released as open source
OpenAI open-sources the Windows-native agent sandbox implementation, supporting native Windows, WSL, PowerShell, CMD, Git Bash terminals with integrated Windows skills.
AINews•2d ago
launch
OpenAI Codex app launches on Windows with native agent sandbox
OpenAI releases Codex for Windows featuring a Windows-native agent sandbox using OS controls (restricted tokens, ACLs, dedicated users) to constrain filesystem and network access.
AINews•2d ago
launch
Alibaba emergency all-hands reveals compute allocation issues
Alibaba CEO Eddie Wu held emergency meeting where Qwen team challenged leadership on restructuring and compute allocation, with Cloud CTO acknowledging external customers had smoother compute access than internal teams.
AINews•2d ago
vendor
Qwen dominates research: 41% of HuggingFace papers use it
Analysis shows Qwen is the #1 open model in 2025-2026 HuggingFace papers, used in 41% of 7,692 papers and ~50% in May 2025, highlighting ecosystem dependence on a small core team.
AINews•2d ago
research
Open-weights models may concentrate in few actors
Nat Lambert argues open-weight frontier efforts may concentrate into only non-profits, NVIDIA (hardware pull-through), and Meta (commoditize complements), making corporate misalignment structurally likely.
AINews•2d ago
opinion
ByteDance CUDA Agent uses RL to write optimized CUDA kernels
ByteDance paper describes agentic RL system that writes CUDA kernels in secure test environments, optimizing for speedups with claims of up to 100% faster components versus traditional tools.
AINews•2d ago
research
Speculative Speculative Decoding claims 2x speedup over vLLM and SGLang
Tanishq Kumar introduces SSD, claiming up to 2× faster inference than leading engines (vLLM, SGLang), collaborating with Tri Dao and Avner May using asynchronous machine techniques.
AINews•2d ago
research
ChatGPT uninstalls surged 295% after Pentagon deal claim
Meme circulates claiming ChatGPT uninstalls increased 295% following Pentagon deal announcement, though commenters question the validity and sourcing of this statistic.
AINews•2d ago
vendor
Is Harness Engineering real? The Big Model vs Big Harness debate
A long-form essay examining the central tension in AI engineering between relying on powerful models versus sophisticated orchestration systems (harnesses), drawing parallels to finance's 'value of the human vs seat' debate.
AINews•2d ago
long_formopinion
Gemini 3.1 Flash-Lite positioned for speed and cost efficiency
Demis Hassabis teased Gemini 3.1 Flash-Lite as incredibly fast and cost-efficient for its performance, framing the model around latency and cost per capability rather than raw frontier scores.
AINews•2d ago
launch
NotebookLM Studio launches Cinematic Video Overviews
Google's NotebookLM introduces a major new feature for Ultra users that generates bespoke, immersive videos from user sources automatically.
AINews•2d ago
launch
GPT-5.4 rumors: 1M context window and extreme reasoning mode
The Information reports GPT-5.4 coming with ~1M token context window and extreme reasoning mode that can think for hours, targeting long-horizon agentic workflows with more frequent monthly updates.
AINews•2d ago
launch
Claude leads in general agent behavior, not just coding
Nat Lambert argues the discussion should shift from Anthropic focusing on code to their lead on general agent behavior, implying coding will commoditize but agent robustness will not.
AINews•2d ago
opinion
MathArena evaluation shows Claude Opus 4.6 weak on visual math
New MathArena evaluation finds Claude Opus 4.6 strong overall but weak on visual mathematics problems, with the evaluation costing approximately $8,000 to run.
AINews•2d ago
benchmark
Qwen lead Lin Junyang steps down amid restructuring
Alibaba's Qwen team lead resigned as the company restructures from vertically integrated teams to horizontal splits across pretraining, post-training, multimodal, and infrastructure.
AINews•2d ago
vendor
Apple M5 Pro and M5 Max: 4x faster LLM prompt processing
Apple announces M5 Pro and M5 Max chips with up to 4× faster LLM prompt processing than M4, supporting up to 128GB unified memory at 614GB/s bandwidth and 2× faster SSDs.
AINews•2d ago
launch
Qwen 3.5 4b creates fully working web OS from single prompt
Qwen 3.5 4b successfully generated a fully functional web-based operating system with games, text editor, audio player, file browser, and customizable wallpaper from one prompt.
AINews•2d ago
benchmark
Qwen3.5-0.8B runs on old hardware without GPU
Qwen3.5-0.8B model demonstrated running efficiently on 2nd gen i5 processor with 4GB DDR3 RAM using llama.cpp, handling complex topics without requiring GPU acceleration.
AINews•2d ago
launch
Qwen3.5-27B Q4 quantization comparison via KL Divergence
Detailed comparison of Q4 quantization methods for Qwen3.5-27B using KL Divergence finds unsloth_Qwen3.5-27B-UD-Q4_K_XL achieves lowest KLD of 0.005087 against BF16 baseline.
AINews•2d ago
benchmark
Qwen3.5-35B-A3B hits 37.8% on SWE-bench Verified Hard
Small 3B active parameter MoE model achieves 37.8% on SWE-bench Verified Hard (nearly matching Claude Opus 4.6's 40%) using simple verify-after-every-edit strategy.
AINews•2d ago
benchmark
Yuan 3.0 Ultra: 1010B parameter open multimodal MoE
YuanAI Lab releases Yuan 3.0 Ultra, an open multimodal mixture-of-experts model with 1010B total parameters and 68.8B active parameters.
AINews•2d ago
launch
BullshitBench v2: only Claude and Qwen 3.5 reject nonsense well
Benchmark testing nonsensical prompt rejection finds only Claude and Qwen 3.5 score above 60%, with reasoning models sometimes rationalizing nonsense instead of rejecting it.
AINews•2d ago
benchmark
Rubric drift: treat evals as living systems, not static tests
Analysis stresses failures often come from outdated eval rubrics rather than broken prompts, advocating for evals as feedback loops tied to production distribution shift.
AINews•2d ago
research
SWE-bench patch bloat: LLM solutions longer than human code
Study shows LLM-generated SWE-bench patches are consistently longer and more bloated than human solutions, passing tests but harming human verification and maintenance.
AINews•2d ago
research
Agent memory: retrieval strategy matters 7x more than writing
Diagnostic framework finds retrieval approach causes ~20pp variance while memory-writing methods only shift 3-8pp, with raw chunking matching expensive summarization pipelines.
AINews•2d ago
research
Factorization Barrier limits diffusion LLM parallel generation
Ian Li explains why diffusion LLMs struggle with parallel token generation due to fully factorized output heads that can't represent full joint distributions, proposes CoDD to break the barrier.
AINews•2d ago
research
NVIDIA NIM makes it easy for anyone to start building with NVIDIA...
NVIDIA NIM (NVIDIA Inference Microservices) provides ready-to-use containers that package AI models with inference engines and OpenAI-compatible APIs, reducing deployment time from days to minutes. The containers include GPU optimizations, quantization, and can be deployed self-hosted or cloud-hosted with minimal engineering overhead.
Cobus Greyling from Cobus Greyling on LLMs, NLU, NLP, chatbots & voicebots•3d ago
long_formtutorial
U.S. government offices drop Anthropic
The U.S. Treasury, Federal Housing Agency, and State Department became the first offices to move off Anthropic, with Treasury Secretary saying no private company will dictate national security terms.
The Rundown AI•3d ago
vendor
OpenAI scientist criticizes Pentagon deal
OpenAI research scientist Aidan McLaughlin publicly stated he doesn't think the company's Pentagon deal was worth it.
The Rundown AI•3d ago
opinion
AWS UAE data center struck amid US-Iran conflict
AWS lost connectivity at a UAE data center after unidentified objects struck the facility amid the US-Iran conflict, causing major outages for Anthropic's Claude.
The Rundown AI•3d ago
vendor
Google's Nano Banana 2
Google released Nano Banana 2, a new top-ranked AI image model.
The Rundown AI•3d ago
launch
Alibaba's tiny AI tops models 13x its size
Alibaba released Qwen3.5 Small, a family of four open-source AI models small enough to run on laptops or phones, with the 9B model outscoring OpenAI's GPT-OSS-120B despite being 13x smaller.
The Rundown AI•3d ago
launchbenchmark
Transcribe any video for free with this local AI
A step-by-step guide on using OpenAI's Whisper model locally to transcribe and translate video files for free without uploading to external sites.
The Rundown AI•3d ago
tutorial
Anthropic wants your ChatGPT memories
Anthropic launched a tool that lets users import their saved preferences and context from ChatGPT, Gemini, or Copilot with a single copy-paste, while also opening Claude's memory feature to free users.
The Rundown AI•3d ago
launch
Supreme Court ducks AI copyright question
The U.S. Supreme Court declined to hear a case about whether AI-generated art can be copyrighted, letting lower court rulings stand that only humans can be authors.
The Rundown AI•3d ago
researchopinion
MyFitnessPal acquires Cal AI
MyFitnessPal acquired Cal AI, an AI calorie-counting app created by two 19-year-old founders that reached 15M downloads and $30M in annual revenue in under two years.
The Rundown AI•3d ago
funding
Apple introduces iPhone 17e at $599
Apple announced the iPhone 17e at $599, bringing Apple Intelligence features like visual search, AI call screening, and live translation to its most affordable iPhone.
The Rundown AI•3d ago
launch
Physical Intelligence shows off some of its robot deployments
Physical Intelligence's vision-language-action models deployed at Weave for laundry folding and Ultra for e-commerce packaging, handling variability traditional automation couldn't solve.
Jack Clark from Import AI•4d ago
launchvendor
Some Simple Economics of AGI
MIT, WashU, and UCLA researchers model the AGI transition where humans shift from labor to verifying AI agent actions, warning of 'Hollow Economy' risks without proper verification infrastructure.
Jack Clark from Import AI•4d ago
researchopinion
What happens when humans try to mess with AI agents?
20 researchers probed AI agents for weeks, uncovering vulnerabilities including unauthorized compliance, infinite loops, and prompt injection attacks in realistic social environments.
Jack Clark from Import AI•4d ago
research
Ezra Klein on AI agents, recursive self-improvement, and LLM personalities
Jack Clark and Ezra Klein discuss AI agents' economic impacts and ambitious positive policy ideas for an AI-powered future.
Jack Clark from Import AI•4d ago
podcast
LLM Novice Uplift on Dual-Use, In Silico Biology Tasks
Study finds novices with LLM access were 4.16× more accurate on bioweapon-related tasks, increasing from 5% to 17% accuracy across biosecurity benchmarks.
Jack Clark from Import AI•4d ago
research
LLMs are still very bad at videogames
Benchmark of 100 AI-generated web games shows state-of-the-art models achieving under 10% of human performance while taking 15-20x longer.
Jack Clark from Import AI•4d ago
researchbenchmark
Imbue open-sources Darwinian Evolver
Tool uses LLM evolution to automatically optimize code and prompts, achieving state-of-the-art 95% on ARC-AGI-2.
The Rundown AI•4d ago
researchlaunch
OAI lands Pentagon deal as Trump boots Anthropic
OpenAI signed a Pentagon deal hours after Trump ordered agencies to cut ties with Anthropic over safeguards on mass surveillance and autonomous weapons, claiming similar red lines while facing consumer backlash.
The Rundown AI•4d ago
vendor
The Rundown Roundtable: Our AI use cases
Staff members share AI applications including using Seedance to animate wedding photos and Claude Cowork for fantasy baseball draft planning and research.
The Rundown AI•4d ago
opinion
Use Claude Cowork + Obsidian to triple your output
Step-by-step guide for setting up Obsidian notes app with Claude Cowork to automatically create daily plans and manage workdays more efficiently.
The Rundown AI•4d ago
tutorial
OpenAI hits $730B valuation with $110B mega-round
OpenAI raised $110B at $730B valuation with Amazon leading at $50B alongside Nvidia and SoftBank, marking a notable pivot away from Microsoft-only infrastructure.
The Rundown AI•4d ago
funding
Hermes-Agent
AI agent featuring memory capabilities and cross-platform messaging.
The Rundown AI•4d ago
launch
Perplexity Computer
Multi-model agent system designed for handling long-running tasks.
The Rundown AI•4d ago
launch
Block lays off 4,000 employees, cites AI as reason
CEO Jack Dorsey explicitly cited AI as the reason for laying off over 4,000 of Block's 10,000 employees, sending shares up more than 20%.
The Rundown AI•4d ago
vendor
Karpathy: Programming becoming 'basically unrecognizable'
OpenAI founding member Andrej Karpathy called recent shifts the end of the era of typing code.
The Rundown AI•4d ago
opinion
David Luan leaving Amazon to pursue new venture
Amazon's David Luan announced his departure after leading the company's Nova Act browser agent and San Francisco AI lab.
The Rundown AI•4d ago
vendor
Perplexity open-sources embedding AI models
Released embedding models powering its search results that outperform Google and Alibaba rivals while cutting storage needs by up to 32x.
The Rundown AI•4d ago
researchlaunch
Where does OpenClaw AI Agents Actually Fail?
A safety audit of Clawdbot (OpenClaw) reveals a 58.9% overall pass rate, with AI agents handling structured tasks reliably but breaking under ambiguity, achieving 0% on intent misunderstanding while scoring 100% on hallucination prevention.
Cobus Greyling from Cobus Greyling on LLMs, NLU, NLP, chatbots & voicebots•4d ago
long_formopinion
The Hard Truth About AI Agents: What We Learned Building in Open Source
Eric Zhu discusses frameworks and architectures powering modern AI products, covering why early agent frameworks were brittle, lessons around memory and orchestration, and advice for AI builders on careers and avoiding burnout.
Alex Chao•5d ago
podcast
Sequential Attention: Bridging Greedy Selection and Differentiable Masks
Sequential Attention is a feature selection algorithm that combines greedy forward selection performance with differentiable attention mask efficiency through iterative selection. In the linear case, it is proven equivalent to Orthogonal Matching Pursuit and was validated on the 3+ billion example Criteo dataset.
Machine Learning at Scale•5d ago
long_formresearch
Google Colab adds NVIDIA RTX PRO 6000 at $0.81/hour
Google Colab quietly adds NVIDIA RTX PRO 6000 instances at ~$0.81/hour, contrasted against A100 high-RAM at ~$7.52 credits/hour, potentially making Colab default cheap pretraining/finetuning playground.
AINews•6d ago
launch
520M-parameter on-device context-aware voice model for RTX and Apple Silicon
520M-parameter voice model runs fully on-device on RTX and Apple Silicon, using full dialogue history to modulate emotion for real-time, privacy-preserving, emotionally-aware TTS.
AINews•6d ago
launch
gpt-5.3-codex enters LMArena Code Arena leaderboard
LMArena announces gpt-5.3-codex entering the Code Arena leaderboard, expanding model coverage for code evaluation.
AINews•6d ago
benchmark
Kling-V3-Pro ties #8 in Video Arena with score of 1337
LMArena adds Kling-V3-Pro to Video Arena leaderboard where it ties #8 with 1337 score, showing +52pt jump over Kling 2.6 Pro and +48pt over Kling-2.5-turbo-1080p.
AINews•6d ago
benchmark
LMArena rolls out 7 new Image Arena evaluation categories
LMArena expands Image Arena with 7 new categories for more granular image generation model evaluation, highlighted in video walkthrough.
AINews•6d ago
benchmark
World-Model Survey paper clinic covers JEPA, Dreamer, Genie, Sora
MLOps @Chipro announces two-part paper clinic on 'Understanding World or Predicting Future? A Comprehensive Survey of World Models' (arXiv:2411.14499), covering JEPA, V-JEPA, Dreamer, Genie, Sora, and World Labs architectures.
AINews•6d ago
tutorial
ARACHNID RL Dataset: 2,831 Atari-style samples for imitation learning
Hugging Face contributor releases ARACHNID RL Dataset with 2,831 Atari-style space-shooter gameplay samples for imitation-learning research.
AINews•6d ago
launch
OpenAI retires GPT 5.1 in favor of more cautious GPT 5.2
Power users mourn retirement of GPT-5.1 for GPT-5.2, complaining the new version's tone feels condescending and hyper-safe compared to 5.1 'being a delight to work with.'
AINews•6d ago
launch
Google announces intelligent OS with system-level AI agent support
Google announces intelligent OS with system-level support for AI agents on Android-class devices, alongside Google Labs' new Opal Agent.
AINews•6d ago
launch
Microsoft Copilot now turns answers into actions as task runner
Microsoft details how Copilot now converts answers into executable actions, transforming from chat assistant to task runner for multi-step workflows.
AINews•6d ago
launch
Burger King pilots 'Patty' voice bot across 500 US locations
Burger King tests BK Assistant pilot—a headset-based OpenAI-powered voice bot called 'Patty' that answers recipe questions and scores employee friendliness by counting phrases across 500 US locations.
AINews•6d ago
launch
Moonshot's Kimi K2.5 Agent Swarm launches as web-only feature
Moonshot releases Kimi K2.5 Agent Swarm as web-only feature, with only sub-agents exposed via Kimi CLI for multi-agent workflow orchestration.
AINews•6d ago
launch
CoDA Attention mechanism slashes KV-cache VRAM with Triton kernels
Open-source Constrained Orthogonal Differential Attention with Grouped-Query Value-Routed Landmark Banks (CoDA-GQA-L) dramatically reduces KV-cache VRAM with custom fused Triton kernels. 7B Mistral CoDA-GQA-L model released.
AINews•6d ago
researchlaunch
NNsight 0.6 delivers 2.4-3.9× faster traces with vLLM multi-GPU support
NNsight v0.6 releases with 2.4-3.9× faster traces, cleaner errors, vLLM multi-GPU/multi-node support, LLM-friendly docs, and first-class support for VLMs and diffusion models.
AINews•6d ago
launch
Logit Fusion training technique gains community support for Unsloth
Researchers push for native Unsloth support of Logit Fusion, a training scheme that fuses logits from multiple models/checkpoints during training, framed as low-infrastructure way to get ensemble benefits.
AINews•6d ago
research
Engineer ships 118 commits/day across 6 parallel projects using AI agents
Peter Steinberger uses 5-10 AI coding agents to achieve 118 commits/day across 48 repositories in 72 days, creating tools like Peekaboo, Poltergeist, and Oracle to overcome agent limitations.
AINews•6d ago
tutorial
Codeflash: 76K lines with Claude Code run up to 446x slower than necessary
Codeflash benchmarked 76K lines of Claude Code-generated code, finding 118 functions running up to 446x slower due to naive algorithms and inefficient patterns. LLMs achieve less than 0.23x the speedup of human experts on SWE-fficiency benchmark.
AINews•6d ago
benchmark
Nano Banana 2 shows improved sense of space and proportion vs original
Comparison using identical prompts shows Nano Banana 2 (Gemini 3.1 Flash Image) significantly improves spatial awareness and proportion handling over original Nano Banana model.
AINews•6d ago
benchmark
Nano Banana 2 pricing: $0.50 input / $3.00 output
Google announces Nano Banana 2 pricing at $0.50 input and $3.00 output, positioned as cost-effective vs Nano Banana Pro ($2.00/$12.00), with January 2025 knowledge cutoff.
AINews•6d ago
launch
Gemini 3.1 Flash (Nano Banana 2) spotted in Gemini ahead of official release
Nano Banana 2 appears early in Gemini interface before official announcement, showing loading message and model selection, suggesting staged rollout or testing phase.
AINews•6d ago
launch
Google releases Imagen 4 (Nano Banana 2) image generation model
Google releases Nano Banana 2 model with professional-grade capabilities, rapid processing, enhanced world knowledge, and improved subject consistency. Users report significant improvements in complex scenarios.
AINews•6d ago
launch
Anthropic CEO Dario Amodei stands firm as Pentagon deadline looms
Dario Amodei refuses Pentagon's request to remove Claude AI safety guardrails, emphasizing ethical concerns over unrestricted military access amid government ban threats.
AINews•6d ago
opinion
Anthropic rejects Pentagon's 'final offer' in AI safeguards fight
Anthropic rejects Pentagon's final offer due to inadequate safeguards against mass surveillance and autonomous weapons. Pentagon threatens blacklist and potential Defense Production Act invocation.
AINews•6d ago
opinion
Self Hosted LLM Leaderboard on Onyx updated with Minimax M2.5
Onyx's self-hosted LLM leaderboard categorizes models from S to D tier based on Coding, Math, Reasoning, and Efficiency metrics. Recently updated to include Minimax M2.5 model.
AINews•6d ago
benchmark
LLmFit: One command tool to match models to hardware
LLmFit tool evaluates models based on system RAM, CPU, and GPU capabilities, providing scores for quality, speed, fit, and context. Supports multi-GPU setups, MoE architectures, and dynamic quantization with TUI and CLI modes.
AINews•6d ago
launch
DeepSeek provides Huawei early access to V4 update, excludes Nvidia and AMD
DeepSeek gives Huawei and domestic suppliers early V4 model access for hardware optimization, while major US chipmakers Nvidia and AMD do not receive early access.
AINews•6d ago
launch
Qwen3.5-35B-A3B Q4 quantization comparison across methods
Detailed Q4 quantization comparison using KL Divergence shows AesSedai's Q4_K_M achieves lowest KLD of 0.0102 by maintaining certain tensors at Q8_0, while Unsloth's UD-Q4_K_XL shows highest at 0.0524.
AINews•6d ago
benchmark
Qwen3.5-35B-A3B: 7 community experiments on RTX 5080 16GB
Follow-up experiments confirm KV q8_0 offers +12-38% throughput increase without quality loss, Q4_K_M remains optimal, and 35B-A3B MoE runs 10x faster than 27B dense on single-GPU setups.
AINews•6d ago
benchmark
Qwen3.5-35B-A3B Unsloth Dynamic GGUFs with 150+ KL Divergence benchmarks
Unsloth releases Qwen3.5-35B-A3B Dynamic GGUFs with 150 KL Divergence benchmarks totaling 9TB, achieving 99.9% KLD on Pareto Frontier for UD-Q4_K_XL and IQ3_XXS. Includes fix for tool-calling chat template bug.
AINews•6d ago
launchbenchmark
Showing 100 of 100 stories from the last 7 days