Model reference · open weights

NousCoder

LLMs NousResearch Text gen 1 build Open weights 480 dl/mo

NousCoder is an open-weight language model from NousResearch. NousCoder-14B (BF16) weighs 29.5 GB; the smallest configuration that runs it is L40S 48 GB.

NousCoder-14B is a 14.8B parameter text-generation model developed by NousResearch and published as NousResearch/NousCoder-14B.
It is a competitive programming model post-trained on Qwen3-14B via reinforcement learning, featuring an 81920 token context length and an Apache-2.0 license.

Summary of the NousResearch/NousCoder-14B model card, 2026-10-01

Hugging Face ↗

What it is

Released by	NousResearch
Released	2025-12-23
Parameters	14.8B
VRAM	29.5 GB for the weights

What it runs on

Memory and cards for NousCoder-14B (BF16)

29.5 GBweights, file size

164 MBcache per 1K tokens

812 MBruntime overhead, at least

81,920 tokenscontext max

Card	Requests at once		Context max	Memory
Card	8K each	32K each	Context max	Memory
RTX 3060 12 GB … RTX 4090 24 GB 4 smaller cards	—	—	—
RTX 5090 32 GB	—	—	3K	31.0 GB
L40S 48 GB	10	2	all 80K	44.0 GB
A100 80 GB	35	8	all 80K	78.2 GB
H100 80 GB	32	8	all 80K	78.1 GB
RTX PRO 6000 Blackwell 96 GB	43	10	all 80K	93.8 GB
DGX Spark (GB10) 128 GB unified	54	13	all 80K	107 GB
H200 141 GB	76	19	all 80K	138 GB
B200 180 GB	104	26	all 80K	176 GB
2× RTX 4090 24 GB tensor parallel	11	2	all 80K	23.4 GB a card
2× RTX 3090 24 GB tensor parallel	11	2	all 80K	23.4 GB a card
2× RTX 5090 32 GB tensor parallel	23	5	all 80K	31.0 GB a card

Memory needed at each load

Requests at once	8K tokens each	32K tokens each
1	31.7 GB	35.7 GB
5	37.1 GB	57.2 GB
8	41.1 GB	73.3 GB
16	51.8 GB	116 GB
32	73.3 GB	202 GB
64	116 GB	374 GB

One card, with vLLM's small-card settings.

From the model card

What NousResearch says about NousCoder

Read the model card

We introduce NousCoder-14B, a competitive programming model post-trained on Qwen3-14B via reinforcement learning. On LiveCodeBench v6 (08/01/2024 - 05/01/2025), we achieve a Pass@1 accuracy of 67.87%, up 7.08% from the baseline Pass@1 accuracy of 60.79% of Qwen3-14B. We trained on 24k verifiable coding problems using 48 B200s over the course of four days.

Acknowledgements

I would like to thank my mentor, Roger Jin, Dakota Mahan, Teknium, and others at the Nous Research team for their invaluable support throughout this project. I would also like to thank Together AI and Agentica for their immensely helpful blog posts on DeepCoder-14B. Finally, thank you to Modal and Lambda for their generous support by providing me with credits.

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

Details

Type	Language models
Task	Text gen
Context	81,920 tokens
Based on	Qwen/Qwen3-14B
Popularity	480 downloads / month
Weights	29.5 GB, NousCoder-14B (BF16, file size)

Trained / evaluated on

livecodebench/code_generation_liteagentica-org/DeepCoder-Preview-DatasetNousResearch/lcb_testNousResearch/RLVR_Coding_Problems