whatmodel.sh / Code Tulu 2 34B

Code Tulu 2 34B

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Code Tulu 2 34B is an open-weights language model from Allen Institute for AI, released November 2023. Its published size is 34B parameters. Licence: AI2 ImpACT Low-risk; Llama 2 base terms.

Catalogue identity notes

Exact full-precision model evaluated in the Tulu 2 paper (arXiv:2311.10702v2). QLoRA ablations are separate and not imported.

Model information sources

Release date ↗ · Weights licence ↗

43.8±1.7#468 of 719 LLMsfrom 5 benchmark results across 1 source

$ how it compares

The LLMs ranked either side of Code Tulu 2 34B, out of 828 with a score, then more from Allen Institute for AI.

#modelscorewhereRAMreleased
555Granite 8B Code Instruct (4K) IBM44±1local~6.6GB2024-04
556Falcon3 10B Instruct TII44±3local~7.6GB2024-12
557Code Tulu 2 34B Allen Institute for AI44±2local~72GB2023-11
558Qwen3 1.7B Alibaba44±1local~2.6GB2025-04
559Phi-4-mini 3.8B Microsoft44±1api $0.13/M +local~3.8GB2025-02
more from Allen Institute for AI
493OLMo 2 13B Instruct (1124) Allen Institute for AI46±1local~9.6GB2024-12
516Tulu 2 70B Allen Institute for AI46±2local~28GB2023-11
599OLMo 2 7B Instruct (1124) Allen Institute for AI42±1local~5.8GB2024-12

see all LLMs · find the ones that fit your machine

$ what it's good at

coding▁▁▁▃▆██▇▆▄▃▁ #310 of 458 * · 17% of its score · 1 test

Beats Solar Open 100B (102B-A12B), GLM-4.7-Flash, Qwen3 4B Instruct 2507 here, though they rank higher overall. Trails DeepSeek LLM 67B Chat, WizardLM 70B V1.0, Phi-4-mini 3.8B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Tulu 2: CodexEval pass@10 lab82.5% ✎▃▃▃·▆▃▃█▃▂▃▅ beats 22/25very easy · avg 86%▇▇▇▇▇▇▇ 17%sharp · at its level
math▁▂▃▅▆▅▄█▄▁▁▁ #295 of 398 * · 14% of its score · 1 test

Beats GPT-4 (0613), Athene-70B, Gemma 2 27B IT here, though they rank higher overall. Trails Gemma 7B (base), DeepSeek R1 Distill Qwen 1.5B, CodeGemma 7B (v1.0) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
GSM8K lab54.0% ✎▆▂▅█▅▅▃▆···▂ beats 17/25easy · avg 73%▇▇▇▇▇▇ 14%sharp · at its level
other tests3 not in a task · 67% of its score
testresultvs the others in its rundifficultycountswhy
BIG-Bench Hard lab64.3% ✎▁·····▂█▂▅·▁ beats 18/25easy · avg 77%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 45%sharp · at its level
MMLU lab53.6% ✎▁▁▃▃▄█▄▁█··▁ beats 14/25easy · avg 74%▇▇▇▇▇▇▇ 16%at its level
Tulu 2: TyDiQA gold passage, 1-shot F1 lab60.6% ✎▂··▁▃▂▂▂█▅▃▂ beats 23/25medium · avg 61%▇▇ 4.9%blunt · at its level

$ where to run it

Other ways to run it

  1. Run it as published · needs ~67.49 GB · from huggingface.co/allenai
    uv run --with torch --with 'transformers<5' --with accelerate python -c "from transformers import pipeline; print(pipeline('text-generation', model='allenai/codetulu-2-34b', torch_dtype='auto', device_map='auto')('Once upon a time', max_new_tokens=40)[0]['generated_text'])"

Weights and documentation: huggingface.co/allenai ↗

$ how the score was worked out · Tulu 2 — base and instruction… 99%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Tulu 2 — base and instruction-tuned evaluations · lab-reported5▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 99%CC BY 4.0
prior · pull toward the middle▇ 1%
weighted average= 43.8

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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