A focused feed of models, agents, research and open-source releases for people building with AI.
Ornith released Ornith-1.5-122B , a 122B MoE reasoning model under MIT license. Fits on a single 80GB GPU; GGUF, FP8, NVFP4, MLX, and Mobile variants shipped day one. Reports 84.5% SWE-bench Verified, 76.9% SWE-bench Pro no-hack, 81.2 Terminal-Bench 2.1.
Prism ML released Ternary-Bonsai-2-2B-gguf , a 2B vision-language model with ternary weights. Language model fits in ~0.52 GB, down from ~3.8 GB at FP16, via PTQ1_0, PQ2_0 or Q2_0 packings. Derived from Qwen3.5-2B with hybrid 3:1 linear/full attention and a 262K-token context window.
VideoZeroBench introduces 500 long-video questions with manually annotated temporal intervals and spatial boxes as evidence. A five-level protocol separates answering from temporal grounding and spatial grounding.
COEVO unifies correctness and PPA optimization into a single evolutionary loop for LLM-based RTL generation . Correctness becomes a continuous fitness dimension, not a binary gate that discards promising buggy candidates.
CCORE GLB-3D-Coder is a 4B Qwen3.5 fine-tune that generates Three.js HTML pages exporting low-poly GLB assets Released Apache 2.0 with Q4_K_M, Q8_0, and Safetensors weights; runs locally via llama.cpp, Ollama, LM Studio, vLLM Training dataset of 3,918 rendered, GLB-validated examples across 40
OpenCode Browse is an MCP server plus Chromium extension for agent-driven real browsers. Agents drive background tabs via Playwright and hand control back for 2FA, logins, or payments. Works with Chrome, Edge, Brave, Arc, Vivaldi, Opera and Chromium family browsers.
Stanford, KAUST, UT Austin and MIT release Muses-Bench , the first systematic multi-user LLM agent benchmark. 26 models tested; GLM-5.3 leads at 90.5 average, Claude-Sonnet-5 tops proprietary at 87.0. Execution accuracy drops sharply when user instructions conflict, even for frontier models.
OpenAI's GPT Image 2.6 Flare and Sunburst take #1 and #2 on both Artificial Analysis image arenas. Flare's median generation time dropped from ~55s to ~30s while matching 2.5 Sunburst quality. Sunburst is slower (~2m15s per image) but wins 7 of 10 text-to-image use cases.
GitHub's new Rubicon storage architecture hits 7,200 pushes per second on one repo. Internal tests sustained 5,000 concurrent agents on a single repository with low read latency. Design splits immutable objects (Azure Blob) from mutable refs (new Ref-CAS log service).
Farbod Tavakkoli published OTel-2.0-LLM-31B-IT-QLoRA , a telecom-tuned Gemma 4 31B variant, on Hugging Face under Apache 2.0. Post-trained on ~440B tokens of 3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN, and TM Forum standards material.
Nous Research released Arete-27B , a Qwen3.8-27B fine-tune reinforcing Western values. Built an 803-prompt eval across 7 categories including free expression, democracy, and privacy. Propaganda rate dropped from 22% to 1% after post-training on ~53M tokens.
Microsoft Foundry adds Kling v3 Pro, LTX-2.5 Fast and Pro, and MiniMax H3 Max and Max Turbo via a fal partnership . Models are offered as a first-party Azure service in public preview, with text-to-video and image-to-video support.
DeepSeek V4.1 Flash on GB300 NVL72 is up to 1.5x faster per agent after a month of stack tuning Single-agent speed jumped from 596 to 848 tok/s; throughput per GPU up ~1.6x at 80 tok/s per agent Gains came from newer vLLM, FlashInfer autotuning, fuller CUDA graphs, tuned spec decoding, and Dynamo
Vals AI released CyberOffense , a benchmark where agents autonomously attack simulated enterprise networks. GPT-6 Sol Daybreak led at 80.8% of objectives, with DeepSeek V4.1 Flash close behind at 72.8%. DeepSeek runs cost under $3 versus around $22 for Daybreak, with no safeguards.
NVIDIA's Lyra 2.0 turns a single image into an explorable 3D world via generative reconstruction. Built on WAN-14B video diffusion; outputs 3D Gaussian splats and meshes ready for Isaac Sim. Fixes spatial forgetting with a per-frame 3D memory cache for information routing.
Frontier Model Forum released Version 1.0 of a Jailbreak Rating Framework for cyber threats. Backed by OpenAI, Anthropic, Google DeepMind, Microsoft, and Amazon security leads. Scores jailbreaks on capability gain, scope, ease of deployment, and ease of discovery.
BioBasicsBench is a new 100-task benchmark from OpenAI and Ligo for wet-lab AI agents. Tasks use raw gel images, colony plates, microscopy, and instrument files across five biology domains. GPT-6 Astra led with 66.7% mean partial-credit reward at high reasoning effort.
Meta rolled out specialized models distilled from Muse Wonder to Muse and Meta AI. Active Memory tracks what context is still relevant versus stale across long conversations. Active Focus helps the model recover, retry, or find alternatives when tasks hit blockers.
ICML paper decomposes the RLHF vs DPO gap into explicit (exact optimization) and implicit (finite sample) representation gaps. No universal winner: RLHF, DPO, or online DPO can each dominate under different model mis-specifications.
Antigravity 2.29 adds Voice Mode powered by the Gemini Live model, rolling out gradually. Supports live two-way conversation, interruptions, multilingual chat, and voice approval of permission prompts. Nine voice options and playback speed adjustable from 0.75x to 2x.
Artificial Analysis launched VIALS-AA , a leaderboard for scientific image reasoning across 161 tasks and 8 domains. Gemini 4 Carbon leads at 64%, GPT-6 Astra 63%, Claude Opus 5.5 61%, measured as pass^3 consistency. Cost per task spans 99x: $0.02 for GPT-6 Luna to $1.77 for Grok 4.7 (xhigh).
Unsloth released FP8/INT8 quants of Qwen-Image-2.1, cutting the transformer from 14.23GB to ~7GB Same-seed LPIPS vs bf16 is 0.064 (INT8) and 0.112 (FP8), with INT8 as the default shipped scheme Dynamic FP8 variant runs on 6GB VRAM via CPU offloading, under 2x slower Base Qwen-Image-2.1 is a 7B
Qwen3-ASR-1.7B is now available as GGUF quantizations for transcribe.cpp local inference. Six variants from BF16 (4.08 GB) down to Q4_K_M (1.32 GB) with WER staying between 1.62% and 1.81%. Supports 52 languages and dialects , including 22 Chinese dialects, Cantonese, and singing voice.
Ai2 released ToyBlocks , an open simulated tabletop for controlled robot learning experiments under Apache 2.0. Ships with 85,336 long-horizon demonstrations, 1,600 hours of execution data, and 144 million frames in LeRobot format.
No stories match these filters.
