<- all tokdocs

Nanbeige4.3-33B, Mesh LLM and Opera WebAI: Local AI News Checked

By , founder of Mavric and Apso

View on TikTok ->

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

Published November 10, 2026. Based on a TikTok by @ai.honeycove. Full video notes (markdown)