Nanbeige Releases a 33B Agent Ahead of a Smaller Model

Nanbeige has released a new AI model for agents that write code, use tools and complete office tasks, while preparing a much smaller version. Its October 25 announcement describes Nanbeige4.3-33B as a teacher for an upcoming three-billion-parameter model. Developers can download the teacher now; the smaller release remains a plan. Announcement

For people choosing hardware, those two versions would be very different purchases. The available model contains roughly 33 billion learned numerical parameters. Its published configuration also allows 524,288 tokens, the pieces of text a model processes. Keeping a conversation that long in memory can require more space than storing the model itself. Configuration

A calculation from its architecture puts an ordinary, uncompressed attention cache at 104 gibibytes for one full-length sequence. That estimate excludes model weights and other working memory. It assumes two-byte cache entries. Shorter conversations, compressed caches and different serving arrangements change the requirement. Configuration

The download history records the weight upload on October 24, followed by configuration and instruction updates. That matters when reproducing results: the published files were still changing ahead of Sunday’s announcement. The latest revision checked for this article is dated October 25. History

A larger teacher for a smaller agent

Nanbeige’s stated strategy starts with a capable teacher, which can help train later models. An open release gives outside developers access to that teacher before the intended compact model arrives. The announcement names Nanbeige4.3-3B as the first planned student, without giving a launch date or its hardware requirements. Announcement

The lab has already published research on a smaller predecessor. Its July report describes Nanbeige4.2-3B, with three billion parameters outside the vocabulary embeddings, trained on 28 trillion tokens. That model sends information through the same stack of layers twice, increasing computation without adding a second set of learned weights. Earlier research

Its researchers also tested sharing the attention cache between those passes. Sharing halved cache memory, they report, but produced smaller performance gains than keeping separate caches. They chose the larger-cache version. That earlier experiment shows a specific tradeoff the team has studied: fitting more computation into a small parameter count can still require extra working memory. The announced 4.3 student has no published architecture here that would establish the same design. Earlier research

What long conversations add to memory

Nanbeige’s weight index lists 66,237,954,048 bytes of tensors. That is about 66.24 gigabytes in decimal units, or 61.69 gibibytes, the binary units used in the cache calculation. Download size and usable device memory need matching units before anyone can judge whether a model fits. These are the original released weights, before any lower-precision conversion. Weight index

Attention lets a language model use earlier text when predicting the next token. A key-value cache stores intermediate information from that earlier text so the model can reuse it. Hugging Face’s cache documentation explains that the default dynamic cache grows as a response continues. A fixed-size cache instead reserves capacity in advance, helping some software optimizations while potentially wasting space on short requests. Cache guide

For this architecture, the calculation uses 52 layers, eight key-value heads per layer and 128 values per head. Storing both keys and values at two bytes each gives 212,992 bytes per token. At 32,768 tokens, that is 6.5 gibibytes; at 131,072, it is 26. The model’s maximum capacity is four times that last length. These figures describe one conventional cache, not a measurement from a running server. Configuration

Adding the full-length cache to the released tensor size gives about 165.69 gibibytes before other allocations. That sum describes the assumptions above, rather than a minimum computer specification: a runtime may divide storage across devices or keep some data elsewhere. It does show why choosing a machine from the roughly 66-gigabyte download alone would miss a potentially larger memory requirement. Weight index

Hugging Face documents cache offloading, which moves most layers’ cached information to system memory and brings it back as needed. Those transfers can reduce generation speed. It also documents lower-precision caches. Each approach changes what stays on the accelerator, so a capacity estimate needs the actual cache method and conversation length alongside the model name. Cache guide

Model quantization addresses another allocation. It stores selected weights with fewer bits per number. Hugging Face’s bitsandbytes documentation describes eight-bit and four-bit alternatives to ordinary linear layers, while some other modules retain their original precision. Its examples also show that weights moved to a CPU can remain in a larger format. A smaller download therefore does not specify every allocation during use. Quantization guide

A team comparing converted downloads would need to establish which parts were compressed, which hardware the conversion supports and whether useful answers survive the change. Those checks are separate from calculating nominal weight size. SnowRock has inspected the release files and reproduced the arithmetic here; it has not run Nanbeige or tested a quantized version. Quantization guide

The card’s sample launch commands split the model across four accelerators, a method called tensor parallelism. They also reserve a specified share of device memory and select parsers for the model’s reasoning and tool calls. These are examples for serving software, rather than a declaration that every deployment needs four GPUs. The appropriate arrangement depends on the weight format, available memory and number of simultaneous requests. Model

Agent scores include the software around a model

Nanbeige reports 80.9 on Terminal-Bench 2.1 and 23.8 on version 4.0, compared with 73.0 and 6.3 for Qwen3.8-27B. Its own table also shows losses: 60.3 against Qwen’s 61.7 on SWE-Bench Pro. Those are developer-reported results. The card identifies Apache 2.0 as the license. Model

The benchmark’s maintainers explain why version numbers matter. Terminal-Bench 4.0 changed computing resources, removed eight tasks and fixed others. Tasks now allow agents eight hours. The maintainers describe these as changes requiring fresh runs. An apparent drop between the two versions cannot, by itself, measure a model getting worse at an unchanged assignment. Benchmark update

A harness is the surrounding software that gives an agent tools and manages its attempts. Nanbeige used different harnesses for different benchmarks, including Terminus 2 and mini-swe-agent. Its comparisons within each benchmark used matching settings. The developers also recommend preserving earlier reasoning during tool-driven conversations, a setting they enabled for evaluation. Model

Nanbeige has not specified when it will release the student. Its architecture and evaluation results would show how much of the larger teacher’s capability survives in the smaller model. Announcement