Black FridayHiggsfield: up to 63% off annual plans - Seedance 2.5, Nano Banana Pro, GPT Image See the deal →affiliate link
Skip to content

promptowy.com

Tencent releases WeMM-Video weights for analysing footage with its original sound

I checked the files behind Tencent’s four WeMM-Video releases. Here is what the downloads contain and how to check an answer against NASA’s Apollo 11 footage.

By: 3 min read
Four WeMM-Video models for analysing picture and sound
Contents 5
  1. When did the weights arrive?
  2. How to check the audio against the picture
  3. Budgeting for a US or UK deployment
  4. Sources
  5. Frequently asked questions

At a glance
  • WeMM-Video can use original audio and frames together to answer in text.
  • The BF16 weights alone come to about 6.57 GB for 2B or 72.36 GB for 35B-A3B.

Is a voice on the Apollo 11 recording coming from the person in the picture or over the radio? That is the sort of question Tencent’s WeChat Vision team has built WeMM-Video to handle. The models take video frames with their original sound and return text. I checked the files behind all four BF16 releases.

When did the weights arrive?

The first complete set, 2B, was committed at 01:41:08 UTC on 20 November. In New York it was still 19 November, at 8:41:08 p.m. The files for the final size in the sequence, 9B, were committed at 05:08:27 UTC, or 12:08:27 a.m. in New York. Those are upload times recorded in the release history. The hosting service does not give a separate timestamp for making a repository public.

How to check the audio against the picture

Tencent supplies an Apollo 11 clip and examples that run it with and without sound. I would start with the audio-only recording and mark who can actually be identified from the voice. Then I would compare the silent picture description and the combined answer. In a descriptive transcript, put visible actions beside the words actually heard. The W3C guide gives examples of the difference, which matters when the speaker is off screen.

The 2B, 4B and 9B variants are dense models. The larger 35B-A3B uses a mixture of experts. I get about 6.57 GB for the 2B weights and 72.36 GB for 35B-A3B, counting only the BF16 weight shards. Config files and runtime memory are separate. Tencent documents both English and Chinese.

NASA Goddard’s archive footage shows Armstrong stepping onto the Moon. Tencent’s example uses footage of the same Apollo 11 sequence to show what changes when sound is available.

Budgeting for a US or UK deployment

The weights carry an Apache 2.0 licence with separate terms for listed third-party components. The release gives no per-minute hosted price in US dollars or pounds sterling. For a hosted deployment, use the US or UK region and currency from the provider’s quote.

Use the file sizes to estimate download time and disk space for a local copy, or compare other AI models before choosing a host.

Sources

Frequently asked questions

Was 20 November the first mention of this project?

The GitHub repository became public on 17 November. The researchers submitted their arXiv paper on 19 November at 16:58:15 UTC. The 20 November release added the complete downloadable BF16 weights.

WeMM-Video produces text. You can ask for a description of a visible action or for an answer that draws on the original audio.

Polish version

Kinetyra DealLovable Pro $119/yrReplit Core $109/yrPostHog $49/yrSee all deals →
× ‹ powiększenie ›