whatmodel.sh / Nanbeige4.3-33B

Nanbeige4.3-33B is an open-weights language model from Nanbeige, released October 2026. Its published size is 33.1B parameters and a 524k-token context window. Licence: apache-2.0.

Catalogue identity notes

Flagship 'teacher' model of the Nanbeige4.3 series (card), post-trained from Nanbeige/Nanbeige4.3-33B-Base (card base_model). HF repo created 2026-10-24 (weights uploaded 2026-10-24; pinned README commit df919e1 on 2026-10-28). Text only; thinking on by default (enable_thinking), reported evals use preserve_thinking=true. Contact address on the card is [email protected]. Licence: apache-2.0 in the card's YAML metadata; the repo has no LICENSE file. No official API and no official Ollama library entry (ollama.com/library/nanbeige and /nanbeige4.3 returned 404 on 2026-11-12). Official card documents Transformers, SGLang and vLLM (tensor parallel 4) with a local MODEL_PATH placeholder, so those are not listed as copy-pasteable routes. Benchmark comparison columns are Nanbeige re-runs of other models and are not used.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

33,118,977,024 BF16 parameters in the repo's safetensors metadata (pinned sha df919e1, observed 2026-11-12); dense LlamaForCausalLM per pinned config.json (52 layers, hidden 7168). The card names it 33B.

No comparable score yet. Its published results are below, but they don't overlap enough with other models to calculate a score.

$ how it compares

Scored LLMs released closest to Nanbeige4.3-33B, out of 1206 with a score, then more from Nanbeige.

#modelscorewhereRAMreleased
3Gemini 4 Carbon Google84±1api $4.00/M2026-10
6GPT-6.1 Terra OpenAI83±1api $2.00/M2026-10
10Muse Wonder Full Meta81±1Muse Code only2026-10
19Kimi K3.1 Pro Moonshot AI78±1api price unknown2026-10
20Muse Wonder Meta78±1api $8.00/M2026-10
more from Nanbeige
474Nanbeige4-3B-Thinking-2511 Nanbeige54±3local~3.7GB2025-11

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasksnot ranked on this task yet · 5 tests
testresultvs the others in its rundifficultycountswhy
Terminal-Bench 2.1 lab80.9% ✎▁·▁·▁▁▄▃▅█▆█ beats 47/68—not counted yet: no other model in the same run to compare it with
SWE-bench Pro (split unspecified) lab60.3% ✎▁·▁▁▁▁▁▄█▁▁▁ beats 46/64—not counted yet: no other model in the same run to compare it with
Terminal-Bench 4.0 lab23.8% ✎▂·▁··▂▄▄▅█▇▄ beats 16/41—not counted yet: no other model in the same run to compare it with
τ³-bench lab48.8% ✎▁▁▂▁·▁▁·▃▁▂█ beats 10/26—not counted yet: no other model in the same run to compare it with
Toolathlon-Verified lab70.4% ✎▂·▂▂▅▃···▃██ beats 9/19—not counted yet: no other model in the same run to compare it with
codingnot ranked on this task yet · 2 tests
testresultvs the others in its rundifficultycountswhy
SWE-bench Pro (split unspecified) lab60.3% ✎▁·▁▁▁▁▁▄█▁▁▁ beats 46/64—not counted yet: no other model in the same run to compare it with
SciCode (level unspecified) lab52.6% ✎▁····▃▃▄▄█▆▃ beats 41/42—not counted yet: no other model in the same run to compare it with
mathnot ranked on this task yet · 1 test
testresultvs the others in its rundifficultycountswhy
AIME 2026 lab95.2% ✎▁▁▁▂▂▂▆█▇▂·▁ beats 41/54—not counted yet: no other model in the same run to compare it with
other tests4 not in a task · 0.0% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond lab89.5% ✎▁▁▁▁▂▃▅▆▇▇█▂ beats 276/336—not counted yet: no other model in the same run to compare it with
NL2Repo lab53.0% ✎▂▂▂·▃▃▅█▃▃▆▃ beats 17/24—not counted yet: no other model in the same run to compare it with
AutomationBench v1.0.6 lab54.2% ✎▂···▂·▂▂··█▆ beats 8/10—not counted yet: no other model in the same run to compare it with
DeepPlanning lab63.3% ✎▂···▂··▄█▄·▃ beats 10/10—not counted yet: no other model in the same run to compare it with

$ where to run it

Memory is what a version takes in GPU memory (or a Mac's unified memory) with an 8k-token conversation. With less, it runs partly from system RAM, more slowly. Enter your machine to see what runs at full speed.

llama.cpp · recommended

version

  1. Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
    macOS / Linux
    curl -LsSf https://llama.app/install.sh | sh
    Windows
    irm https://llama.app/install.ps1 | iex

    other ways to install ↗

  2. Run Nanbeige4.3-33B · needs 23.3 GB memory, 20.8 GB download · from huggingface.co/bartowski
    chat in the terminal
    llama cli -hf bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q4_K_M

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run Nanbeige4.3-33B · needs 23.3 GB memory, 20.8 GB download · from huggingface.co/bartowski

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF@Q4_K_M -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

  1. Install Ollama

    download it (macOS / Windows) ↗

    Linux
    curl -fsSL https://ollama.com/install.sh | sh
  2. Run Nanbeige4.3-33B · needs 23.3 GB memory, 20.8 GB download · from huggingface.co/bartowski
    chat in the terminal
    ollama run hf.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q4_K_M

Model repository and documentation: huggingface.co/Nanbeige ↗

$ published results · 11 results, not enough overlap for a score
testresultsource and conditions
Terminal-Bench 2.180.9%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'Terminal-Bench 2.1' (Agentic Coding) = 80.9; footnote 2: 'Terminal-Bench 2.1 is evaluated using Terminus 2.'

Terminal-Bench 4.023.8%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'Terminal-Bench 4.0' (Agentic Coding) = 23.8; footnote 3: 'Terminal-Bench 4 and DeepSWE are evaluated using mini-swe-agent v2.4.6.'

SWE-bench Pro (split unspecified)60.3%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'SWE-Bench Pro' (Agentic Coding) = 60.3; footnote 3 marker: mini-swe-agent v2.4.6. Public/commercial split not stated, so not mapped to swe-bench-pro.

NL2Repo53.0%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'NL2Repo' (Agentic Coding) = 53.0; footnote 5: 'NL2Repo is evaluated using Claude Code.'

SciCode (level unspecified)52.6%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'Scicode' (Agentic Coding) = 52.6; main-vs-subproblem level not stated.

AutomationBench v1.0.654.2%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'Automation Bench' (Agentic Cowork) = 54.2; footnote 6: 'Automation Bench v1.0.6 includes 600 publicly released tasks.'

Toolathlon-Verified70.4%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'Toolathlon-Verified' (Agentic Cowork) = 70.4; footnote 10: 'Toolathlon-Verified sets no max_tokens limit for individual model requests.'

DeepPlanning63.3%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'DeepPlanning' (Agentic Cowork) = 63.3; footnote 12: evaluated using the official repository.

τ³-bench48.8%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'Tau3-Bench' (Agentic Cowork) = 48.8; footnote 12: evaluated using the official repository. Domain average as printed.

GPQA Diamond89.5%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'GPQA Diamond' (Reasoning) = 89.5.

AIME 202695.2%Nanbeige/Nanbeige4.3-33B model card · maker-reported

Nanbeige own evaluation, pinned Nanbeige/Nanbeige4.3-33B card (sha df919e1), section '2. Model Performance' table, column 'Nanbeige4.3-33B' (footnote 1: evaluated with preserve_thinking=true in the chat template). Card: 'All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.' Comparison columns (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) not used. Row 'AIME 2026' (Reasoning) = 95.2; I/II split and sampling count not stated.

$ improve this page

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json