limoce

> ollama gpt-oss 120b mxfp4 1 94.67 11.66

This is insanely slow given its 200+GB/s memory bandwidth. As a comparison, I've tested GPT OSS 120B on Strix Halo and it obtains 420tps prefill and >40tps decode.

limoce

11 months agodiscuss

"Train a 100,000,000,000B multimodal large language model" got 75/100

Submissions

World Tour Finals 2026 Algorithm: Human vs. AI Exhibition Match

China's DeepSeek developing its own AI chip, sources say

China Blocks Meta's $2B Acquisition of AI Firm Manus

Claude Code Source Map

NeurIPS retracts overly broad sanctions list

NeurIPS Statement on Sanctions

Declining to review for NeurIPS due to policy on sanctioned institutions

PetClaw: A Desktop Pet Companion That Works 24/7

Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search

Cursor's 'Composer 2' model is apparently just Kimi K2.5 with RL fine-tuning

Add NPU support for AICore for Gemma4 model

SteptronOss: Lightweight, AI-native training framework for large language models

[untitled]

Adding Support for Qwen3.5

Reduce Complexity of Matmul to O(N^2logN) by Asymptotically Optimal Quantum Algo

Learning from context is harder than we thought

Residual Context Diffusion

Step 3.5 Flash

LongCat-Flash-Lite 100B A3B Technical Report [pdf]

Jinja.cpp: A single-header C++11 Jinja2 template engine for LLM chat templates

comments