Stable Code 3B

+ contribute

Stable Code 3B is an open-weights language model from Stability AI, released January 2024. Its published size is 2.8B parameters and a 16k-token context window. Licence: stability-ai-community.

Catalogue identity notes

Maker repo stabilityai/stable-code-3b @5cee3fa: licence tag other. 2,795,443,200 parameters from the Hugging Face safetensors metadata (dense, so active = total). Released 2024-01: month the maker's Hugging Face repo was created (createdAt 2024-01-09). Base completion model; not stable-code-instruct-3b. Ollama tags stable-code:3b-code-q4_K_M, stable-code:3b-code-q8_0 (sizes as listed on the tags page, observed 2026-09-29).

32.7±0.6#621 of 719 LLMsfrom 15 benchmark results across 4 sources

$ how it compares

The LLMs ranked either side of Stable Code 3B, out of 828 with a score, then more from Stability AI.

#	model	score	where	RAM	released
716	Vicuna 7B v1.5 LMSYS	33±1	local	~5.8GB	2023-08
717	Guanaco 33B University of Washington	33±2	local	~70GB	2023-05
718	Stable Code 3B Stability AI	33±1	local	~3.3GB	2024-01
719	OLMo 2 1B Instruct (0425) Allen Institute for AI	33±2	local	~2.4GB	2025-04
720	DeepSeek-Coder 1.3B Base DeepSeek	32±1	local	~2.4GB	2023-10
	more from Stability AI
746	StableLM-3B-4E1T Stability AI	29±1	local	~3.3GB	2023-09
750	StableLM-Base-Alpha 7B v2 Stability AI	29±1	local	~30GB	2023-08
785	StableLM-Base-Alpha 3B v2 Stability AI	24±1	local	~13GB	2023-08

see all LLMs · find the ones that fit your machine

$ what it's good at

coding▁▁▁▃▆██▇▆▄▃▁ #423 of 458 · 92% of its score · 10 tests

Beats DeepSeek R1 Distill Qwen 7B, Qwen3.5 2B, DeepSeek R1 Distill Llama 8B here, though they rank higher overall. Trails Llama 3.2 1B Instruct, SmolLM2 1.7B Instruct, Mistral 7B Instruct v0.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
HumanEval +2	29.3% ✎	▁·▁▁▃▄▆▅█▃▁▁ beats 14/99	easy · avg 79%	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 48%	sharp · at its level · several runs
HumanEval+	25.6% ✎	▁·▁▁▁▃▄▄▄█▃▁ beats 14/99	easy · avg 74%	▇▇▇▇▇▇▇ 17%	sharp · at its level
MBPP +1	34.8% ✎	█·▅▃▃▃▅▃▃▅·▅ beats 4/15	easy · avg 82%	▇▇▇ 6.9%	at its level · several runs
HumanEvalSynthesize (HumanEvalPack, 6-language average)	27.8% ✎	▁·▂▁▃▆▅█▄▄▁▂ beats 5/35	medium · avg 64%	▇▇ 4.8%	at its level
MBPP+	45.8% ✎	▁···▁▁▂▅▇█▆▂ beats 12/75	easy · avg 68%	▇▇ 3.8%	at its level
DS-1000	22.7% ✎	▄▄·▄██▄█·█·▄ beats 4/12	medium · avg 52%	▇ 2.8%	typical
HumanEvalExplain (HumanEvalPack, 6-language average)	12.3% ✎	▄▇█▃▄▇▄▁▇▃·▁ beats 3/35	medium · avg 49%	▇ 2.4%	typical
CRUXEval-I	33.5% ✎	▃▅█▃▃▅·▃▃▅·▅ beats 6/15	medium · avg 59%	▇ 2.3%	at its level

2 more tests

test	result	vs the others in its run	difficulty	counts	why
HumanEvalFix (HumanEvalPack, 6-language average)	10.5% ✎	▁·▁▁▄██▄█▁▅▅ beats 7/35	medium · avg 56%	▇ 1.9%	sharp
CRUXEval-O	32.1% ✎	▃·▃██▅▃·▃▃▅▃ beats 4/15	medium · avg 55%	▇ 1.9%	blunt · at its level

math▁▂▃▅▆▅▄█▄▁▁▁ #356 of 398 * · 2.0% of its score · 1 test

Beats Yi 34B Chat, Llama 3 8B Instruct, Command R (03-2024) here, though they rank higher overall. Trails Gemma 3 1B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
GSM8K	39.7% ✎	▄▄▄·▄█·██·█▄ beats 7/12	easy · avg 75%	▇ 2.0%	sharp · at its level

other tests1 not in a task · 6.3% of its score

test	result	vs the others in its run	difficulty	counts	why
EvoEval (overall)	20.7% ✎	▂▂▃▄▄█▁▄▅▃▁▃ beats 6/42	medium · avg 55%	▇▇▇ 6.3%	typical

$ where to run it

llama.cpp · recommended

version

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 ↗
Run Stable Code 3B · needs ~1.7 GB · from huggingface.co/TheBloke
chat in the terminal
llama cli -hf TheBloke/stable-code-3b-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf TheBloke/stable-code-3b-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Stable Code 3B · needs ~1.7 GB · from huggingface.co/TheBloke
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/TheBloke/stable-code-3b-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Stable Code 3B · needs ~1.7 GB · from ollama.com
chat in the terminal
ollama run stable-code:3b-code-q4_K_M

MLX (Mac)

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Stable Code 3B · needs ~1.8 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/stable-code-3b-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/stable-code-3b-4bit

Official weights and documentation: huggingface.co/stabilityai ↗

$ how the score was worked out · EvalPlus / EvalPerf published… 47%, Top Leaderboard Ranking = Top… 20%, StarCoder 2 and The Stack v2:… 17% and 1 more

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

source	results	share of the score	licence
EvalPlus / EvalPerf published evaluation runs	3	▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 47%	Apache-2.0
Top Leaderboard Ranking = Top Coding Proficiency, Always? EvoEval: Evolving Coding Benchmarks via LLM — authors' own runs	2	▇▇▇▇▇▇ 20%	CC BY 4.0
StarCoder 2 and The Stack v2: The Next Generation — authors' own runs	4	▇▇▇▇▇ 17%	CC BY-SA 4.0
Granite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runs	6	▇▇▇▇▇ 16%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 32.7

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit 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 2026-10-11 02:20 UTC · sources and licences · json