Nanbeige built a 33B "teacher" first, so tomorrow's small agents can learn from it 🧑‍🏫

Most model launches are the end of a story: here is the model, here are the scores, go use it. The release Nanbeige put on Hugging Face this weekend reads more like the first chapter. Nanbeige4.3-33B is an open-weight model under Apache 2.0, and the team describes it plainly as the "flagship teacher model" for its 4.3 series. The plan is to build a strong agent at around 30B parameters first, then distil its agentic skills into much smaller models, starting with an upcoming Nanbeige4.3-3B.

I like that framing a lot, because it puts distillation front and centre instead of hiding it in a footnote.

What's actually in the box

The model card and config are refreshingly concrete:

  • A dense model, no mixture-of-experts tricks. The config shows a Llama-compatible architecture with 52 layers, a hidden size of 7,168, 56 attention heads and 8 key-value heads (grouped-query attention).
  • 512K context. The card lists support for up to 524,288 tokens.
  • Built for tools. The chat template has an enable_thinking switch (on by default) and a preserve_thinking option that keeps earlier reasoning in multi-turn tool use. The team recommends keeping that on for code-agent and office workflows, and XML-style tool calls for the best results.
  • A Base model too. Nanbeige4.3-33B-Base sits next to the instruct model, which is great news for anyone who wants to fine-tune from scratch.

It runs with Transformers, SGLang or vLLM.

The numbers, with an honest caveat

Nanbeige compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all at their highest thinking effort and with the same harness. On its own table it leads most agentic benchmarks: 80.9 on Terminal-Bench 2.1 (versus 73.0 for Qwen3.8-27B), 65.5 on DeepSWE and 70.4 on Toolathlon-Verified. It doesn't win everywhere. Qwen3.8-27B is slightly ahead on SWE-Bench Pro (61.7 versus 60.3) and GDPval, and Gemma4-31B leads GPQA Diamond.

These are self-reported scores, so I'd treat them as a promising starting point until independent evaluations come in. The candid touch of showing where the model loses makes me trust the table a bit more, though.

Why this is the exciting part

The community response says a lot. Within about two days, community members had already uploaded dozens of derived repos to the Hugging Face hub: GGUF builds for llama.cpp, MLX versions for Apple Silicon, EXL3, FP8 and NVFP4 quantisations, and more. A permissive license plus a plain Llama-style architecture means people can run it on their own hardware almost immediately. For a homelab, a 4-bit 33B model is very much within reach.

The bigger idea is the teacher-student pipeline itself. If a strong 33B agent can pass its planning and tool-use habits down to a 3B model, capable agents could run on laptops, phones and school devices, without every step going to the cloud. That's exactly the direction I hope open AI keeps taking, and I'm curious to see how much of the 33B's agentic skill survives the trip down to 3B.

Sources

A teacher model released in the open is a gift to everyone who wants to train smaller ones. I'll be watching for that 3B student. 🚀

Reacties

Een reactie posten

Populaire posts van deze blog

SpaceX just asked to fly Starship hundreds of times a year 🚀

Last week the U.S. Federal Aviation Administration took a quiet but important step: it opened a formal review of SpaceX's request to launch Starship from Starbase in South Texas up to 456 times per year . Today the site is cleared for only 25 flights a year. That jump is the clearest signal yet that Starship is meant to become routine hardware, not a rare test rocket. What the FAA is reviewing On 14 October 2026 the FAA published a Notice of Intent to prepare a Programmatic Environmental Impact Statement (PEIS) for SpaceX's Starbase plans. A PEIS looks at a whole program of operations rather than a single launch, which is exactly what you need if you want a reusable rocket to fly the way airlines fly planes. SpaceX's proposal comes in two phases, according to the Federal Register notice: Phase 1: up to 100 launches and 200 landings per year (100 Ship, 100 Super Heavy), any day or time of week. Phase 2: up to 456 launches and 912 landings per year at Starbase. La...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

Google's Deep Review wants to be the second pair of eyes on every long document 🔍

AI has made writing fast. Checking has not kept up. That gap is exactly what Google Research is going after with Deep Review, which went public on 14 October. It's an agentic system that reads a long technical document and tries to find what's wrong with it: broken proofs, logic gaps, unsupported claims. For anyone working on legal and document AI, like me, this is a launch worth a close look. From conference pilot to public tool Deep Review is the grown-up version of Google's Paper Assistant Tool (PAT). Over the past year, authors of more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027 opted in for private feedback before the submission deadline, fully separate from official peer review. The need is real: Google counts over 74,000 submissions to ICLR, NeurIPS and ICML this year, up from about 23,000 in 2023. Now anyone can use it. Every user gets two free reviews per week, and you can connect a Gemini API key for higher limits and programmati...
Meer lezen