AI News
Daily AI industry briefings - model releases, research, infrastructure, and pricing - summarized neutrally from vetted sources, with every claim linked to its origin.
- Latest
Meituan releases 1.6T-parameter agent model trained entirely on Huawei chips; Gemini 4 Carbon tops math benchmark
Meituan open-sourced LongCat-2.5, a 1.6-trillion-parameter agent model trained without Nvidia GPUs, while Google's Gemini 4 Carbon claimed the MathArena benchmark lead despite significantly higher.
The Decoder MarkTechPostRead → -
Stanford Open-Sources Memorilla Memory Module; Runway Releases Open-Weight Robotics Model Praxis-1
Stanford researchers released Memorilla, a memory compression system for frozen LLMs; Runway open-sourced Praxis-1 robotics model under Apache 2.0 license.
MarkTechPost The DecoderRead → -
Google Releases TPU Sync for KV Cache Transfer; Gemini 4 Hydrogen Arrives with 1M-Token Window
Google released TPU Sync, an open-source library reducing KV transfer latency by over 90% on large models, while launching Gemini 4 Hydrogen, a faster agentic model with expanded context and lower.
MarkTechPost The Decoder Hugging Face BlogRead → -
ElevenLabs Releases Scribe v3 Speech Models; Google Launches Gemini Decisions API
ElevenLabs released Scribe v3 speech recognition models with improved accuracy in noisy environments, while Google introduced a specialized decision model API for high-volume classification tasks at.
The Decoder MarkTechPostRead → -
Anthropic ports Claude to AMD in 3 days using AI; open models advance on-device inference
Anthropic used its Mythos model to port Claude to AMD GPUs in three days with minimal human intervention, while open-weight releases target efficiency gains for on-device deployment.
The Decoder MarkTechPostRead → -
Anthropic releases Claude Fable 5.2 and Mythos 5.2; Meta launches Muse Ember for agent tasks
Anthropic released Claude Fable 5.2 and Mythos 5.2 with dual safeguard levels and improved coding performance; Meta released Muse Ember, a small 1M-context model optimized for agent work at lower.
MarkTechPost The Decoder Google DeepMind TechCrunchRead → -
Baidu's ERNIE Evolve Beats Benchmark Records; Google Expands TPU Lineup with Encryption and Density Variants
Baidu's automated research system outperformed open benchmarks while Google shipped two new TPU 8 variants targeting security and compute density.
MarkTechPost The DecoderRead → -
Real-World Coding Agent Benchmark Reveals 74% Human-Cleanup Rate; Google Releases Spec-First Agent Plugin
Alibaba's Co-Design Bench shows even top AI coding agents fail to fully solve three out of four real-world tasks; Google AI releases Elephant-Goldfish plugin for spec-driven code review.
The Decoder MarkTechPostRead → -
Kinference Control Plane cuts TTFT 95%; Ripple-65B achieves top open-source coding benchmark
Buoyant releases Kubernetes inference control plane with KV-cache-aware load balancing; Recursive Intelligence open-sources 65B agentic coding model.
MarkTechPost The DecoderRead → -
Together AI's Agent-Driven Compiler Boosts Inference 2.6x; Box Deploys Nemotron for Token-Efficient RAG
Together AI released an agent-based compiler that continuously re-optimizes inference pipelines, achieving up to 2.6x throughput gains. Box integrated Nvidia's Nemotron for retrieval with 90% fewer.
MarkTechPost TechCrunchRead → -
Anthropic Releases FlashProve for Cryptographic Model Verification; Tencent Releases Unified Vision Model
Anthropic released FlashProve, a zkSNARK system proving AI responses originate from committed model weights without revealing parameters. Tencent released WeVisOne-VL, a 2B/4B open-weight vision.
MarkTechPostRead → -
Anthropic's Message Threads cut TTFT by 35% for long conversations; Zyphra releases hierarchical attention for subquadratic scaling
Anthropic optimizes long-context inference latency via incremental message transmission; Zyphra demonstrates sparse attention retrofitting on frozen models.
The Decoder MarkTechPostRead → -
DeepSeek and Nace AI Release Open-Weight Agent and Decision Models; OpenAI Shifts to Token-Based Pricing
DeepSeek open-sourced dsh-superagent, a router-worker coding agent that cuts p95 task completion time by 47%; Nace AI released Drex Max v1, a 27B multimodal decision model for local Apple Silicon.
MarkTechPost The DecoderRead → -
Factory Droid 1.0 Coding Agent Harness; Vercel Labs Open-Sources agent-desktop for Computer Use
Factory released Droid 1.0, a model-independent coding agent harness achieving 65.2% on Terminal-Bench 4; Vercel Labs open-sourced agent-desktop for cross-platform agent computer use.
MarkTechPostRead → -
GLM 5.3 Autonomously Built Firefox Exploits; Mistral Releases Trillion-Parameter Model Weights
Open-weight GLM 5.3 generated working Firefox exploit chains without human input in under 48 hours for ~$1,000 per exploit; Mistral released Large 4 weights, a trillion-parameter coding model.
The Decoder TechCrunch MarkTechPostRead → -
Nvidia releases open-weight speech model for 40 languages; IBM's Granite Switch enables model updates without retraining
Nvidia released Nemotron 3.5 Transcribe, an open-weight speech LLM with 4.3B active parameters for multilingual transcription; IBM's Granite Switch 4.1.1 allows model updates via instruction layer.
MarkTechPost Hugging Face BlogRead → -
FlashMap Achieves 6.14x LLM Decoding Speedup; Liquid AI Releases 2B Multimodal Models
Polymathic AI and Princeton researchers released FlashMap, a speculative decoding method that drafts 16 tokens in one pass for up to 6.14x lossless speedup on small models. Liquid AI released six.
MarkTechPost Hugging Face BlogRead → -
Opera Launches WebAI Standard for Browser-Based Local Inference; Mirendil Releases Fern 0.1 with 10M-Token Context
Opera introduces WebAI standard for local AI deployment in browsers; Mirendil releases Fern 0.1 model pretrained with 10M-token context window for agentic AI work.
TechCrunch MarkTechPost The DecoderRead → -
Tsinghua's AKV cuts LLM agent inference costs 27.7%; TokenGate accelerates long-context prefill 11.3x
Tsinghua researchers open-sourced AKV, a KV cache optimization server that reduces agent inference costs by 27.7% on average; Stanford and UC Berkeley researchers released TokenGate token pruning.
MarkTechPostRead → -
Nvidia open-sources Flora for robot video labeling at 1/125th GPT-6 Astra cost; Microsoft releases decentralized agent orchestration framework
Nvidia's Flora data engine cuts robot video labeling costs from $88 to $0.72 per hour using fine-tuned open models; Microsoft open-sources Agensh for scaling agent swarms to 1,024 workers.
The Decoder MarkTechPostRead → -
Anthropic Open-Sources Sandbox-Runtime for Agent Isolation; Moonshot Kimi K3.1 Pro Reaches 1M Context
Anthropic releases OS-level sandboxing for AI agents without containers; Moonshot's open-weight MoE model scores near Claude Opus 5.5 with 1M token context.
MarkTechPost The DecoderRead → -
Anthropic's Claude generates formal proofs for open-source code; Cohere ships Rerank 5 with passage-level retrieval focus
Anthropic demonstrated Claude generating 180 million lines of formal proofs for 123 libraries, catching bugs missed by testing; Cohere released Rerank 5 with context-aware passage extraction and.
The Decoder MarkTechPostRead → -
IBM Granite Speech 5.0 achieves top ASR performance; Recursive releases Flint 40B MoE coding model
IBM released Granite Speech 5.0 2B with 5.68% English WER ranking second on Open ASR Leaderboard; Recursive open-sourced Flint, a 40B MoE coding model with 2.5B active parameters reporting 74.2% on.
MarkTechPostRead → -
Deepseek open-sources DeepInfer inference engine; Google Cloud releases llm-d 1.0 for distributed LLM serving
Deepseek open-sourced its inference framework powering V4.1-Flash; Google Cloud released llm-d 1.0, a Kubernetes-native distributed inference stack for TPU and GPU fleets.
The Decoder MarkTechPostRead → -
Cloudflare's Clairic and Inception's Mercury 3 advance extraction and inference speed
Cloudflare released Clairic, a specialized extraction model at $0.10/M input tokens; Inception released Mercury 3 diffusion LLM running at 1,068 tokens per second.
MarkTechPost The DecoderRead → -
MiniMax releases M3.1 Flash 428B MoE at $0.15/M tokens; NVIDIA open-sources Run:ai GPU orchestration
MiniMax shipped a 428B-parameter mixture-of-experts agentic model with competitive token pricing and 1M context window; NVIDIA released Run:ai GPU orchestration as open source.
MarkTechPostRead → -
Google open-sources Harness2 agent framework; Voxel51 adds MCP tooling for multimodal data
Google released Harness2, a test-time agent refinement framework that improved GLM-5 by 11.5 points without weight updates; Voxel51 3.0 unified multimodal AI data platform with 45+ MCP tools.
MarkTechPostRead → -
DeepSeek V4.1-Flash-1017 cuts agentic costs; StepFun releases 600B MoE Step 5 Preview
DeepSeek released V4.1-Flash-1017 with improved agentic benchmarks and lower per-task costs; StepFun shipped Step 5 Preview open-weight MoE with 1M context window.
The Decoder MarkTechPostRead → -
Meta's Muse Spark Open; NVIDIA Nemotron 3.5 Super VL; Black Forest Labs FLUX 3 Image Klein
Meta released an 816B open-weight MoE model optimized for Apple silicon; NVIDIA and Black Forest Labs shipped enterprise agent and image generation models for on-device deployment.
MarkTechPost The DecoderRead → -
Fish Audio releases 35B open-weight MoE text-to-speech model; TypeSafe AI expands Jev with images and 262k context
Fish Audio released Drama 3, an open-weight mixture-of-experts text-to-speech model with 3B active parameters covering 83 languages; TypeSafe AI expanded Jev 1.14 with image input and extended.
MarkTechPostRead →
Summarized from vetted sources, every claim linked. For information only.