First screen, captured 9 days ago. Open the video.

Tencent Hy4

What HN thinks

Another open-weight model lands, and the thread immediately debates geopolitics, coffee lawsuits and whether $2.50 per million tokens is a war crime.

China's open science is refreshing
Some praise Tencent for open-sourcing science models, contrasting it with US safety caution; others say it's competition, not altruism, and export rules may end it.
Price and benchmarks are mixed
Commenters see it as mid-pack and slow but on the cost/accuracy frontier per task, though per-token pricing looks worse than DeepSeek.
Cyber capabilities filtered out
Users note offensive security training was removed, which some call nerfing and others call a reasonable choice.

Consensus: Most agree Hy4 is a solid open-weight model that still trails top proprietary models, and its value depends heavily on cost per task.

Recommended in the comments

GLM (product): Suggested for tasks like security work that Hy4 refuses.
DeepSeek (product): Much cheaper per token, especially cached input, for many-turn coding work.
Artificial Analysis ↗ (service): Cited for comparing cost per task rather than per token.

Summarised by Claude from the comment thread 8 days ago.

Points: 252
Comments: 90
Best rank: #11
Posted: 2026-11-04 by Philpax
Category: AI & ML

tencent-hy hy4 llm open-source-models multimodal computer-use benchmarks

All captures:

Full page

Captured 9 days ago. Click for the whole screenshot.

Looks similar

Screenshot of blog.samaltman.com: Successful People (2013)

Successful People (2013)(blog.samaltman.com)

13 points by epestr 9 days ago | 7 comments

A short 2013 blog post by Sam Altman arguing that the most successful founders pursue a religion-like mission rather than setting out to build companies.

sam-altman founders startups mission-driven-companies

Screenshot of lesswrong.com: I struggle to believe that AI R&D is ingenious

I struggle to believe that AI R&D is ingenious (lesswrong.com)

136 points by porridgeraisin 7 days ago | 155 comments

A LessWrong shortform comment by Arjun Panickssery arguing AI agents show poor research judgment in autonomous R&D side projects like a Scrabble engine and philosophy research.

ai-research-agents autonomous-r-and-d llm-capabilities lesswrong

Screenshot of github.com: Huawei releases MACaQ, its internal quantizer for local inference

Huawei releases MACaQ, its internal quantizer for local inference (github.com)

121 points by thefox96 2 days ago | 30 comments

GitHub repository for Huawei CSL's MACaQ, a model-agnostic quantization tool for text, image, video and audio models targeting llama.cpp, vLLM, Diffusers and PyTorch.

quantization gguf llama-cpp vllm

Screenshot of huggingface.co: Kimi K3.1 Pro model weights are out

Kimi K3.1 Pro model weights are out (huggingface.co)

607 points by abrarthought 9 days ago | 174 comments

Hugging Face model card for Moonshot AI's Kimi K3.1 Pro, an open-weight 2.8T-parameter multimodal MoE model with 1M-token context, including architecture and benchmark results.

kimi moonshot-ai open-weights llm

Screenshot of github.com: 6 Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here

6 Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here (github.com)

56 points by samirswr 4 days ago | 9 comments

GitHub repo for a small LLM inference engine built from scratch in C++ and CUDA, with a YouTube video series and 80+ pages of notes.

llm-inference cuda cpp kv-cache

Screenshot of github.com: Show HN: Reflota – Open-source data recovery for NTFS/FAT disks with phased copy

Show HN: Reflota – Open-source data recovery for NTFS/FAT disks with phased copy (github.com)

15 points by christ_dev 13 hours ago | 2 comments

GitHub repository for Reflota, an open-source, read-only Windows tool for recovering deleted and lost files from NTFS, FAT and exFAT disks, with phased disk imaging.

data-recovery ntfs fat32 exfat