Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8

GitHubNov 7, 7:16 AM PT

Read the original at github.com ↗Hacker News ↗ · 77 pts · 14 comments

Source: GitHub (github.com)
Published: Nov 7, 7:16 AM PT
First seen here: Nov 7, 6:26 PM PT
Edition: Saturday, November 7, 2026, ranked 7
Status: The company itself announced or confirmed this.
Engagement: As shown by each platform when we last looked (Nov 8, 6:27 PM PT).
Why it’s here: Open-source model compression work shared on Show HN.
Topics: AI, Open source

Launch

Muse Ember: Our smallest, most efficient model (opens research.meta.ai)

Meta's entry-level model, described by SiliconANGLE as costing significantly less than the competition and the first trained using Muse Wonder.

research.meta.ai9:58 AM PTNov 9, 2026 Details

News

DeepSeek says its open-source agent app Harness hit 1.54M DAUs on November 4, less than three months after launch (opens bloomberg.com)

DeepSeek says Harness reached 1.54 million daily users within three months of launch and processes about 92 trillion tokens a day, roughly 60 million per user.

Bloomberg12:10 PM PTNov 8, 2026 Details

DiscussionUnverified claim

Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU (opens github.com)

GitHub2:22 AM PTNov 8, 2026 Details

Launch

Thinking Machines Lab debuts Inkling 2 Small, a MoE model with 594B total and 17B active parameters, and plans to release weights next week (opens thinkingmachines.ai)

Artificial Analysis says it matches Gemini 3.8 Flash (high) on its index and would be the top open model outside China once weights ship.

thinkingmachines.ai4:15 PM PTNov 5, 2026 Details

Launch

Tencent Launches Hy4, One of the Leading Open-Source AI Models (opens tencent.com)

tencent.com12:21 PM PTNov 5, 2026 Details

Launch

Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights) (opens linum.ai)

My brother and I run a 2-person lab that trains our own generative video models from scratch. The ultimate goal is a set of creative tools that make it easier for anyone to make animation. Video models today are too slow/expensive and lack sufficient controls. Creativity is an iterative process.

linum.ai8:45 AM PTNov 5, 2026 Details

Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8 · Keeping Up with Silicon Valley

Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8

Related stories

Muse Ember: Our smallest, most efficient model (opens research.meta.ai)

DeepSeek says its open-source agent app Harness hit 1.54M DAUs on November 4, less than three months after launch (opens bloomberg.com)

Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU (opens github.com)

Thinking Machines Lab debuts Inkling 2 Small, a MoE model with 594B total and 17B active parameters, and plans to release weights next week (opens thinkingmachines.ai)

Tencent Launches Hy4, One of the Leading Open-Source AI Models (opens tencent.com)

Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights) (opens linum.ai)