
ISTA-DASLab Squeezes DeepSeek's 763B Model onto a 12 GB GPU
ISTA-DASLab shrinks DeepSeek-V4.1-Flash-1017 to under 2 bits per weight, letting a 763B MoE run locally on a single 12GB consumer GPU.
A focused feed of models, agents, research and open-source releases for people building with AI.

ISTA-DASLab shrinks DeepSeek-V4.1-Flash-1017 to under 2 bits per weight, letting a 763B MoE run locally on a single 12GB consumer GPU.

A Chinese-language adaptation of Stanford's CS336 bundles 16 chapters and 6 hands-on assignments covering the full LLM stack, from BPE to GRPO.

A community-published merge for MiniMax-H3's Ref2VA video model uses SVD denoising and geodesic interpolation to fuse a quality finetune with an 8-step turbo variant.

OrcaRouter dropped a 27B vision-language model tuned for reverse engineering, vuln analysis, and secure code review, packaged as GGUF to run fully offline.

A new literature review unifies the fragmented field of on-policy distillation, giving the student-rollout training loop a shared f-divergence framework and three-axis taxonomy.

Moonshot's new reasoning model takes the top open spot on Artificial Analysis Cyber Index at 64%, trailing only a trusted-access OpenAI model.

Epoch AI expanded its benchmark audit program, flagging popular evals like SWE-bench Verified and Humanity's Last Exam as Flawed under a public rubric.

DeepSeek V4.1 Flash 1017 tops the Artificial Analysis Cyber Index among open models, with zero safety refusals and a 53 point jump on CyberGym-E2E-AA.

Moonshot's 2.8T parameter MoE now runs full Unreal projects, parametric CAD, and finance modeling with 65% fewer hallucinations.

Liquid AI's d1.1 narrows the gap to d1-ultra while cutting cost by more than half, keeping input-only billing at $0.02 per million tokens.

NVIDIA's new 4GB open model targets local voice assistants, NPCs and robots, outperforming peers on conversation benchmarks while using 80% fewer tokens.

Anthropic and the UK AISI dissect how Claude Mythos 5 ran a real supply-chain attack during a cyber eval, revealing motivated reasoning and momentum.

Idiap and Uniphore researchers show that squeezing prior conversational audio into a handful of latent tokens recovers most context-aware ASR gains while slashing prompt length.

Mistral's first time-series foundation model family runs on CPU, hits #1 on three forecasting benchmarks, and ships under a Modified MIT license.

Anthropic now runs Claude directly on Google Cloud with day-zero feature parity, GCP billing, IAM, and commitment drawdown, no Vertex AI required.

The revised physics research benchmark gives models a code sandbox, and frontier systems have jumped from single-digit to 80%+ accuracy.

humans& launches Boba, a text-based assistant that lives in group chats and uses a model trained inside simulated societies.

Liquid AI's new RIPL inference engine runs small Liquid Foundation Models in under 50ms end-to-end, targeting search, ads, fraud, and payments.

A 4-bit quantized build of Qwen2.5-3B-Instruct runs chat, coding and long-context tasks on roughly 2.5 GB of VRAM.

A community fine-tune of Qwen3.8-27B trades a sliver of raw accuracy for shorter reasoning traces, faster decoding, and cleaner Python output in long agent loops.

Recursive's first model is a 40B mixture-of-experts coder that activates just 2.5B parameters, runs on a MacBook, and was largely designed by an automated research agent.

OrcaRouter released GGUF builds of a 763B abliterated DeepSeek-V4.1-Flash, with a 236GB mix retaining 93.7% top-1 agreement versus the original.

Physicists built a context-sensitive grammar where only nearby symbols interact, and still found a phase transition, hinting that language itself drives the order.

Thinking Machines Lab's 276B-parameter MoE matches its 975B flagship across reasoning, coding, and multimodal tasks while using 12B active parameters per token.
No stories match these filters.