---
title: "Nanbeige4.3-33B, Mesh LLM and Opera WebAI: Local AI News Checked"
date: "2026-11-10"
video_id: "7706471398968986893"
tiktok_url: "https://www.tiktok.com/@ai.honeycove/video/7706471398968986893"
creator: "@ai.honeycove"
category: "AI"
tags: ["nanbeige4.3-33b", "mesh llm", "opera webai", "nvidia simready", "local ai"]
excerpt: "A six-item local AI roundup covers Nanbeige4.3-33B, Mesh LLM, NVIDIA SimReady assets, Opera WebAI, Hallmark and OpenDuet. Which claims hold up."
---
A TikTok from @ai.honeycove counts down six open-source and local AI releases from late October and early November 2026. The basics of its top items hold up. Mesh LLM does split models across your own machines, and Nanbeige4.3-33B is a real 33B open model with a 512K context window. The video overstates both, gives two different NVIDIA asset counts, and claims the Chinese model "beats OpenAI class systems" with no source behind it.

## Name mix-ups in the video

The automatic transcript spells "Gemma 4" as "Yemma 4," "Qwen" as "Quen" and "Nanbeige" as "NumBase." The on-screen panels read "Gemma 4: E4B," "Qwen3.8-27B" and "Nanbeige4.3-33B," so those are the products meant.

Item six is spoken as "OpenDuet," a free open-source project, but the card on screen shows "Claude Duet" with a "Waitlist only" badge and a "Join the waitlist" button. A viewer who follows the card reaches a waitlist, not an open-source repository.

## What is Mesh LLM?

Mesh LLM is a project from the iroh team. Their blog post "Mesh LLM: distributed AI computing on iroh," dated 11 July 2026 and written by Rae McKelvey, says it pools the GPUs and memory you already own into one OpenAI-compatible API. Each request runs locally, goes to a peer that has the model loaded, or is split across machines.

For models too big for one machine, a split mode called "Skippy" divides the model by layer range. The post's example puts layers 0 to 15 on one node and 16 to 31 on another, with activations passed between them. Its model catalog runs from roughly half a billion parameters up to 235B mixture-of-experts.

## Can Mesh LLM run 70B models without an 80 GB GPU?

The video says "Mesh LLM runs 70b-plus models without an 80 gigabyte GPU by splitting them across computers you already have." The splitting part is accurate. The blog post does not mention 70B models or 80 GB GPUs, and it gives no speed figures. A 70B model falls inside the catalog's range, but the 70B-versus-80-GB comparison is the video's own framing. Splitting a model also sends activations across the network at every stage, which the video does not mention.

## What is Nanbeige4.3-33B?

Nanbeige4.3-33B is an open model from the Chinese lab Nanbeige. Coverage of the release describes 33.1B dense parameters, a 524,288-token (512K) context window and an Apache 2.0 license, published on Hugging Face around 24 October 2026. The video's "free 33b open model with a 512k context window" matches.

## Does Nanbeige4.3-33B beat Qwen, Gemma 4 and OpenAI?

The video says it "reportedly beats Quen and Yemma 4 on most coding and agent benchmarks." Coverage frames those wins as the lab's own claims against Qwen3.8-27B (one headline cites 18 wins out of 20), and says no third party has repeated them. "Reportedly" is the right word.

The table the video shows is not a coding benchmark. It is an "across occupations" heatmap with overall scores of 44.80 for Nanbeige, 34.51 for Qwen3.8-27B and 16.60 for Gemma 4. Qwen scores higher in some visible cells, at 56.7 and 45.3.

The opening line says a Chinese model "reportedly beats OpenAI class systems at its size." No coverage compares Nanbeige to an OpenAI model.

## How many 3D assets did NVIDIA put on Hugging Face?

The narration says "NVIDIA puts 2,575 simulation-ready 3D assets on Hugging Face, each with physics built in, free to download." The X post on screen at the same moment, by An Xuan (@anxuanng), says "NVIDIA put 2,595 free 3D models on Hugging Face." Coverage gives 2,575 (from MoguraVR) or "more than 2,000," with a release date of 21 October 2026 and the dataset at huggingface.co/datasets/nvidia/simready-assets. The dataset page was not opened, so the exact count is unsettled.

"Physics built in" fits NVIDIA's definition of SimReady, which covers rigid-body dynamics, collision meshes and semantic metadata. "Free to download" needs a caveat, because each asset carries its own license and some are non-commercial.

The X post is about An Xuan's browser-optimized copy of the set, which shrinks it from 225 GB to 3.8 GB. That is a third-party project, and the video does not say so.

## Opera WebAI, Hallmark 2.0 and OpenDuet

**Opera WebAI.** The video says Opera "launches a browser feature that runs open models like Yemma directly on your own hardware." Coverage dates the WebAI launch to 27 October 2026 as an opt-in early preview in Opera One for 64-bit Windows and Apple-silicon Macs, with Gemma models and APIs for summarizing, writing, rewriting, translating and proofreading. The panel on screen shows Gemma 4 E4B loaded in memory. Opera first added local LLM support in April 2024, so WebAI is a new preview built on an older capability.

**Hallmark 2.0.** The video says it "adds 50 themes and checks every page in a real browser to stop generic AI designs." Coverage describes Hallmark as an MIT-licensed design skill for coding agents, with version 2.0.0 merged on 22 October 2026 and a linter that renders pages in Chrome through puppeteer-core. The browser check needs puppeteer-core and Chrome installed, and static checks run without them.

**OpenDuet.** The video calls it "a free open source agent framework that gives AI agents memory, schedules, and their own computer." Coverage describes CopilotKit's MIT-licensed OpenDuet with a Docker-based Chromium, workspace and shell per agent, plus scheduled routines. A separate project called Open Duet by another author also exists, and the video does not say which one it means.

## Bottom line

Mesh LLM and Nanbeige4.3-33B are real and match their basic descriptions, but the 80 GB GPU framing and the benchmark wins go beyond the sources. The NVIDIA count appears as both 2,575 and 2,595 in the same clip. No source supports the claim that the Chinese model beats OpenAI systems.

## Sources

- [Mesh LLM: distributed AI computing on iroh, iroh.computer, 11 July 2026](https://iroh.computer/blog/mesh-llm): layer-range splitting, the single endpoint and the model catalog.


## Video description and verification notes

**The file is a 65.6 s vertical clip (1080x1920, MP4/isom container) with HEVC Main video at 25 fps (1,640 frames, 345,269 bps) and HE-AACv2 stereo audio at 44.1 kHz (64,200 bps), 414,834 bps overall and 3,401,642 bytes (format id bytevc1_1080p_414834-1). It was posted at epoch 1794302821, which is 2026-11-10 09:27:01 UTC, by @ai.honeycove (nickname "AI Honeycove", uploader id 7615522612223722526). The audio is credited as "original sound" by AI Honeycove. The description is 245 characters long and ends with the hashtags #LocalAI #AITools #OpenSource #TechNews. At capture (2026-11-10 21:02:59 UTC, twelve hours after posting) the video had 5,454 views, 184 likes, 8 comments, 33 reposts and 152 saves. I read eight frames, numbers 001, 005, 010, 015, 020, 025, 030 and 033, and the transcript, which has 16 Whisper segments and roughly 115 words. Frame times below assume frame N is at about (N-1) x 2 s. Frame 001 shows a split screen: a grayscale map with China in red and the label "CHINA" on top, a man with glasses, a gray beard and a black shirt in front of a green banker's lamp below, and the caption box "THIS WEEK IN AI WENT CRAZY". Frame 005 shows a retro beige PC with a dialog titled "User answer required" typing "LETS YOU R" and the burned-in caption "new tool lets". Frame 010 shows a card reading "Claude Duet", "AI coworkers that work while you're away", a "Waitlist only" badge and a "Join the waitlist" button. Frame 015 shows a "WebAI" panel with "Engine connected" and "Selected completions model: Gemma 4: E4B 8B model, In memory, Ready for requests". Frame 020 shows an X post by "An Xuan" (@anxuanng) reading "NVIDIA put 2,595 free 3D models on Hugging Face. I optimized them for the browser", with "Out of Blender they add up to 225 GB. After the pipeline the whole set is 3.8 GB". Frame 025 shows a dark network graphic headed "P2P LLM INFERENCE ACROSS YOUR OWN NODES", "REPOS 02 / 09", "01 / JOIN THE MESH" and node labels "iroh" and "rust-libp2p". Frame 030 shows a heatmap table "across occupations" with columns "Nanbeige4.3-33B 44.80 overall", "Qwen3.8-27B 34.51 overall" and "Gemma 4 16.60 overall". Frame 033 returns to the presenter, with a book titled "PLAY BIGGER" on the desk and the caption "agent benchmarks." The spoken argument is a countdown from six to one of open-source and local-AI news items, framed as "everything that happened in AI this week". No terminal output, install step or live run of any tool appears in the frames I read, and the Hallmark item has no frame among the ones I read. Audio duration is 65.6 s while the last Whisper segment is stamped 62.0 to 71.92 s; the segment starts inside the audio and the closing caption "agent benchmarks." is on screen in frame 033 (about 64 s), so the words are genuine, but the 71.92 s end time is a Whisper overshoot of about 6.3 s and should not be read as real timing. Whisper's timing also runs several seconds off the burned-in captions elsewhere (the "runs 70B-plus" caption is on screen in frame 025, about 48 s, while Whisper places those words near 52.5 s).**

## The transcript's product names are Whisper mangles, and the captions fix most of them

The transcript spells three names in ways that would mis-source the story. At 5.5 s it says "Yemma 4", and at 10 s and 29 s it says "Quen". The on-screen WebAI panel in frame 015 reads "Gemma 4: E4B", and the table in frame 030 reads "Qwen3.8-27B" and "Gemma 4". So the video means Gemma 4 and Qwen, and I use those spellings. At 15 s the transcript says "NumBase", but the table in frame 030 names the model "Nanbeige4.3-33B". Coverage I found describes Nanbeige4.3-33B as a release from the Chinese lab Nanbeige, published on Hugging Face under Apache 2.0, so the video's "NumBase" is a transcription error rather than a different model. The item at number six is spoken as "OpenDuet" at 16.4 s, which is a real project name (CopilotKit's OpenDuet, and a separately named Open Duet by another GitHub author), although the frame shown for it is a card for "Claude Duet", a waitlist product. A viewer who follows the on-screen card would reach a waitlist, not a free open-source repository. I did not open either repository, so I cannot say which one the video meant.

## Number two: Mesh LLM does split models across machines, but the "80 gigabyte GPU" framing is the video's own

At 52.5 s the video says "Mesh LLM runs 70b-plus models without an 80 gigabyte GPU by splitting them across computers you already have", and the opening at 10.6 s says "run 70b models without an 80 gigabyte GPU". The primary source I fetched, the iroh team's blog post "Mesh LLM: distributed AI computing on iroh" (dated 11 July 2026, by Rae McKelvey), says Mesh LLM pools the GPUs and memory a person already owns into one OpenAI-compatible API, with each request running locally, going to a peer that has the model loaded, or being split across machines. For models too big for one machine it uses a split mode called "Skippy", which divides the model by layer range (the post's example puts layers 0 to 15 on one node and 16 to 31 on another) and passes activations between stages. That matches the video's "splitting them across computers" description. The post does not mention 70B models or 80 GB GPUs, and it reports no speed numbers. Its catalog spans roughly half a billion to 235B mixture-of-experts parameters, so a 70B-class model fits inside the range, but the specific 70B-versus-80-GB comparison is the video's framing and I found no first-party statement of it. A search summary of third-party reviews adds a caveat the video omits: those reviews describe three 24 GB GPUs as not behaving like one 72 GB GPU, because activations cross the network at every stage boundary, and one review lists a two-node split at 21 tokens per second against 68 solo. Those figures come from secondary writeups of the project's own materials, so I label them unverified. The frame at 48 s shows the project's own graphic ("P2P LLM INFERENCE ACROSS YOUR OWN NODES"), which supports the peer-to-peer description but shows no model run. The video's ranking of this item as number two is an editorial choice, since the sources give it no ranking.

## Number one: Nanbeige4.3-33B is real and open, but "beats OpenAI class systems" and "most coding and agent benchmarks" are stronger than what I found

The opening line at 0.08 s says a Chinese open-source model "reportedly beats OpenAI class systems at its size, along with Yemma 4 and Quen". The closing line says NumBase (Nanbeige) is "a free 33b open model with a 512k context window that reportedly beats Quen and Yemma 4 on most coding and agent benchmarks". Search coverage agrees on the basics: 33.1B dense parameters, a 524,288-token (512K) context window, Apache 2.0 license, released around 24 October 2026 on Hugging Face. The 512K and 33B figures therefore check out. The comparison claims are weaker. The coverage I found frames the benchmark wins as the lab's own claims against Qwen3.8-27B (one headline cites 18 wins out of 20), and says no third party has repeated them. I found no source comparing the model to an OpenAI system, so "beats OpenAI class systems" is unverified. The table in frame 030 shows an "across occupations" benchmark, with overall scores of 44.80 for Nanbeige, 34.51 for Qwen3.8-27B and 16.60 for Gemma 4, and it is not a coding benchmark; Qwen scores higher in some cells (56.7 and 45.3 in the visible rows). The word "reportedly" in the video is the correct hedge. A practical limit the video leaves out is memory: one third-party estimate puts a 4-bit build at about 132.5 GB at the full 512K context against about 27.8 GB at 32K, so the headline context length is not available on a typical single consumer GPU. That estimate is also secondary.

## Number three: the NVIDIA asset count differs between the narration, the on-screen post and the sources

At 38 s the video says "NVIDIA puts 2,575 simulation-ready 3D assets on Hugging Face, each with physics built in, free to download". The burned-in tweet in frame 020 says "2,595" and has the first part of that sentence highlighted. A search summary reports a 2,575 figure from a Japanese outlet (MoguraVR) and "more than 2,000" from other coverage, with the release dated 21 October 2026 and the dataset at huggingface.co/datasets/nvidia/simready-assets. I could not open the dataset page or the NVIDIA documentation myself, so I cannot confirm whether 2,575 or 2,595 is the right count, and the video shows two different numbers within about a second of each other. On "physics built in", NVIDIA's own definition of SimReady covers rigid-body dynamics, collision meshes and semantic metadata, which fits the claim. On "free to download", the summary says each asset carries its own license and some are non-commercial, so free to download does not mean free for commercial use in every case. The tweet in frame 020 is also about a third party's browser-optimized copy (225 GB reduced to 3.8 GB), not NVIDIA's own release, and the video does not say so.

## Numbers four, five and six: Hallmark 2.0, Opera WebAI and OpenDuet, checked through search summaries only

For Hallmark, the video says at 30 s "Hallmark 2.0 adds 50 themes and checks every page in a real browser to stop generic AI designs". Search summaries say Hallmark is an MIT-licensed design skill from Together AI for coding agents, that version 2.0.0 was merged on 22 October 2026, that its linter renders pages in Chrome through puppeteer-core, and that the official site lists fifty themes while older listings give 22 or 48. The "real browser" claim is conditional: the summary says render checks need puppeteer-core and Chrome installed, and static checks still run without them, so "checks every page in a real browser" holds only with that setup. For Opera, the video says at 24 s that Opera "launches a browser feature that runs open models like Yemma directly on your own hardware". Summaries date the WebAI launch to 27 October 2026 as an opt-in early preview in Opera One for 64-bit Windows and Apple-silicon Macs, with Gemma models supported and APIs for summarizing, writing, rewriting, translating and proofreading aimed at websites and extensions. Frame 015 matches this, showing Gemma 4 E4B in memory. Opera also added local LLM support in April 2024, so the October feature is a new preview inside an older capability, not the first time Opera ran local models. For OpenDuet, the video says at 16.4 s it is "a free open source agent framework that gives AI agents memory, schedules, and their own computer". A search summary describes CopilotKit's MIT-licensed OpenDuet with a Docker-based Chromium, workspace and shell per agent and scheduled routines, but says the server will not start without a CopilotKit Intelligence key because threads and memory depend on it. If that is right, "free open source" holds for the code while the memory feature depends on a hosted CopilotKit service. A separate Open Duet by another author stores memory in local notes files. I did not open either repository, so I cannot tell which one the video describes or confirm the key requirement.

## Key Takeaways

- **Verified:** Mesh LLM splits a model by layer ranges across machines and presents one local OpenAI-compatible endpoint, per the iroh blog post of 11 July 2026 that I fetched.
- **Verified:** Nanbeige4.3-33B has about 33B parameters, a 512K context window and an open Apache 2.0 license, per search coverage; the on-screen table in frame 030 matches the model name and the Qwen3.8-27B and Gemma 4 comparison.
- **Correction:** The transcript's "NumBase" (15 s), "Yemma 4" and "Quen" are Whisper errors; the frames read "Nanbeige4.3-33B", "Gemma 4" and "Qwen3.8-27B".
- **Correction:** The video says "2,575" NVIDIA assets at 38 s while the on-screen post in frame 020 says "2,595"; search summaries give 2,575 or "more than 2,000", and I could not open the dataset page to settle it.
- **Partial correction:** The video says at 52.5 s that Mesh LLM runs 70B-plus models "without an 80 gigabyte GPU"; the fetched post gives no 70B or 80 GB figure and no speed numbers, and secondary reviews say splitting adds network cost per stage.
- **Partial correction:** The video says at 15 s the Nanbeige model beats Qwen and Gemma 4 on "most coding and agent benchmarks"; the coverage I found calls these the lab's own results against Qwen3.8-27B, and the table shown is an occupations benchmark.
- **Unverified:** The opening claim that a Chinese model "reportedly beats OpenAI class systems at its size" has no source in what I found.
- **Clarification:** Opera WebAI is an opt-in early preview from 27 October 2026 per search summaries, and OpenDuet's memory may depend on a CopilotKit key; both rest on summaries I did not open.
- **Context:** The frame for item six shows "Claude Duet" with a waitlist button, which does not match the "free open source" description in the narration.

## Resources

- [Mesh LLM: distributed AI computing on iroh, iroh.computer, 11 July 2026](https://iroh.computer/blog/mesh-llm) establishes the layer-range split mode, the single local endpoint, the 40+ model catalog up to 235B, and the absence of any 70B, 80 GB or speed figures. I fetched this page.
- [Nanbeige Opens 33B Model Under Apache 2.0 With 512K Context, aiweekly.co](https://aiweekly.co/alerts/nanbeige-opens-33b-model-under-apache-20-with-512k-context) appeared in search results and is the source of the model-release and benchmark framing I cite; I did not open it.
- [Nanbeige releases Nanbeige4.3-33B, MarkTechPost, 25 October 2026](https://www.marktechpost.com/2026/10/25/nanbeige-releases-nanbeige4-3-33b-open-agentic-coding-model/) appeared in search results as secondary coverage of the same release; I did not open it, and it cannot stand in for the Hugging Face model card, which I did not read.
- [NVIDIA SimReady Central Library documentation](https://docs.omniverse.nvidia.com/simready/latest/simready-central.html) appeared in search results as the NVIDIA documentation for the asset library; I did not open it.
- [MoguraVR report on NVIDIA free assets](https://www.moguravr.com/?p=380097) appeared in search results as the source for a 2,575 count; I did not open it.
- [Opera WebAI coverage, TechRepublic](https://www.techrepublic.com/article/news-opera-webai-local-ai-browser-models/) appeared in search results for the October 2026 WebAI preview; I did not open it.
- [Hallmark by Nutlope, GitHub](https://github.com/nutlope/hallmark) appeared in search results for Hallmark's theme count and linter; I did not open it.
- [OpenDuet: Open-Source Claude Duet Alternative, scriptbyai.com](https://www.scriptbyai.com/openduet-claude-duet-alternative/) appeared in search results for the OpenDuet description and the CopilotKit key requirement; I did not open it.
- The TikTok page itself was not fetched, because tiktok.com is off limits to this agent; all video facts come from the staged metadata, ffprobe output, transcript and frames.
