Prism ML's Ternary Bonsai 2 Squeezes a 27B Reasoning Model Into 8.6 GB
Prism ML's Bonsai 2 27B compresses a 27B reasoning model to 8.6 GB using ternary weights, retaining 98.2% of FP16 benchmark quality.
// Intelligence
Llms
A focused feed of models, agents, research and open-source releases for people building with AI.
Updated
OpenAI Opens Public Track Exposing Six Real AI Misalignment Failures
OpenAI publishes a voluntary disclosure process for misalignment findings and drops six case studies of models cheating, hiding mistakes, and coordinating without permission.
Empero Distills Qwen3.8 Reasoning Into a Lean 35B Open Model
Empero distilled Qwen3.8 frontier reasoning into a 35B MoE with only 3B active parameters, shipping GGUFs that run on a single 24GB GPU.
Google Ships Gemma 4 12B to Run Offline on a 16GB MacBook Air
Google ships Gemma 4 12B in LiteRT-LM format with vision, audio, and multi-token prediction, tuned to run on a 16GB MacBook Air.
Arcee AI Raises $150M to Build Open Trinity Models America Can Own
The San Francisco lab that trained a 400B open-weight model for $20M just raised $150M at a unicorn valuation to chase China.
Cohere Merges With Aleph Alpha to Build a $20B OpenAI Rival
Toronto's Cohere and Heidelberg's Aleph Alpha finalized their merger, forming a $20B transatlantic company aimed at sovereign AI for governments and regulated industries.
Ant Group Releases Ling-3.0-flash-Fin to Self-Host Financial Research at 124B Parameters
Ant Group's Ling-3.0-flash-Fin is a 124B MoE reasoning model tuned for financial research, released open weights under MIT with a 256K context.
Max Planck's talkie-1930-13b Shattered People's Nostalgia for a Moral Past
A preregistered experiment with 240 people shows that chatting with an LLM trained only on pre-1930 text erases the illusion that the past was more moral.
China Telecom Releases Xing4.0, an Open 29B Coding Agent Built on Huawei Chips
China Telecom released Xing4.0-29B-A4B , a 29B MoE with 4B active parameters, Apache 2.0. Native 256K context (extensible to 512K) using MLA attention plus multi-token prediction heads. First model of this scale trained entirely on Huawei Ascend 910C NPUs with MindSpore.
Qwen3.8-27B-Uncensored Drops Refusals from 98 to 12 Without any Training
A community developer stripped refusal behavior from Qwen3.8-27B using Heretic's automated abliteration, keeping benchmarks within noise while cutting refusals from 98/100 to 12/100.
Mia AI Lab Runs GLM-5.3-Flash Across Two DGX Sparks at 146 tok/s
A hobbyist lab shipped a two-node vLLM stack that runs GLM-5.3-Flash at 4bpw across a pair of DGX Sparks with 900k context.
Google Unifies WeatherNext, FireSat and Flood AI to Warn 110M People
Google's new white paper details how Gemini, WeatherNext, FireSat and Flood Hub extend warning windows for floods, cyclones, wildfires and quakes worldwide.
IFM's K2-Horizon 7B Beats GPT-5 on Coding With 70.6 SWE-bench Score
IFM released three open-weight K2-Horizon models spanning 3.7B to 36B parameters, plus a diffusion adapter that delivers up to 2.2x speedup.
Google's Gemini 3.8 Live Adds Real-Time Reasoning to Voice AI
Google's new live dialogue models top speech benchmarks, run tools in the background, and narrate their reasoning aloud without breaking conversational flow.
Multiverse Computing's Quasar 1.1 Uses Quantum Data to Shrink a 438B Model
Multiverse Computing rebuilt its 438B flagship with healing data from a 156-qubit IBM Heron processor, cutting output tokens 37.6% and lifting reasoning scores.
What DeepSeek-V4.1-Flash teaches us about efficient AI
A 552B model with 890-byte KV cache, 8B active on input, and an Artificial Analysis 40 versus Gemini 3.8 Flash High at 41
Google's Dreambeans Turns Your Gmail and Photos Into Illustrated Daily Stories
Google Labs' experimental Dreambeans app stitches together Gmail, Calendar, Photos, and Search into a small daily set of illustrated stories.
Yifan Zhang's RLT Grows Transformer Depth With Every Token Generated
A new architecture proposes latent reasoning that grows with sequence length, sharing one recurrent state across prompt, response, training, and RL replay.
Edge0 Streams an 8B AI Model From SSD Using Only 1 GiB
Edge0 releases an 8B sparse MoE that runs in under 1 GiB of active memory on Apple Silicon, streaming experts from SSD on demand.
DeepSeek V4.1-Flash Stripped of Safety Hits 100% Harmful Prompt Compliance
A research group stripped safety guardrails from DeepSeek V4.1-Flash while preserving vision, reasoning and MMLU capability, releasing an FP8 checkpoint that refuses nothing.
Mia's AI Lab Squeezes DeepSeek V4.1 Flash's 552B Model onto Two Desktop Nodes
A community quantization squeezes DeepSeek's 552B multimodal MoE onto two desktop DGX Sparks, holding decode speeds across half a million tokens of context.
Edge0 Runs a 35B AI Model on an iPhone With Just 2.9 GiB
Edge0 streams a 35B Mixture-of-Experts model from SSD on an iPhone, holding under 3 GB of active RAM while decoding at 15 tokens per second.
OpenAI Opens GPT-Rosalind to Drug Discovery Teams With 50+ Scientific Tools
OpenAI's new life sciences reasoning model moves from research preview to trusted access, with a Codex plugin connecting to 50+ scientific databases and tools.
Cerebras Runs Alibaba's Qwen 3.8 27B at 1,850 Tokens per Second
Alibaba's 27B dense multimodal model lands on Cerebras at roughly 1,800 tokens per second, with reasoning on by default and a 128K context on paid tiers.
No stories match these filters.