granite-4.1-3b

3.40B parameters, text in, text out, 128K tokens of context.

Verified 2026-10-25 Rank #464 in LLMs · score 77.6 method v1 Sources 3 Downloads 30d 608.9k · #150 of 619 tracked

Can I run it?2.0GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24

Do people use it?608.9kdownloads, 30 days#150 of 619 we track · top 24%+4 likes, downloads −5% this week118 likesHugging Face2026-10-24

Worth reading

Picked by rule from 3 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 328 HN pts

Granite 4.1: IBM's 8B Model Matching 32B MoE

IBM released Granite 4.1, a family of open source models trained on 15 trillion tokens with a rigorous data quality pipeline.

The 3B model extends context only to 128K tokens, not the 512K available to larger variants.
Phase 5 extended the context window eventually to 512K tokens for the 8B and 30B models.

firethering.com 2026-04-30 328 HN pts · 209 comments 1.9k words about the granite 4.1 family family

Latest · 6 months ago

The IBM Granite 4.1 family of models

IBM released the Granite 4.1 family of models under an Apache 2.0 license, offering language models in 3B, 8B, and 30B parameter sizes alongside vision, speech, embedding, and guardian capabilities.

Granite Speech 4.1 2B NAR generates entire sequences at once instead of one token at a time.
The Granite 4.1 8B instruct model matches or outperforms the Granite 4.0 32B Mixture-of-Experts model.

research.ibm.com 2026-04-30 88 HN pts · 16 comments 1.7k words about the granite 4.1 family family

3 best of 3 tracked

The 2 picks above, plus these 1. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleGranite 4.1 LLMs: How They're Builthuggingface.co · 2026-04-29 · 1 HN pts

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 66 quantized builds and 3 Spaces.

As a service0no API provider listed
On own hardware66quantized builds
Built on4122 fine-tunes · 19 adapters · 0 merges
In apps3Hugging Face Spaces

Top Spaces running it: ibm-granite/granite-guardian ♥ 15

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 2.0 GB at INT4, 4.1 GB at FP8, 8.2 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

3 dated articles and 14 repo commits so far

Teal is the weights (2026-04-06), amber triangles are 3 dated articles (each links to it), grey ticks are repo commits.

What people built on it

107 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

107 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF ibm-granite/granite-4.1-3b-GGUF for llama.cpp, Ollama, LM Studio 268.9k downloads MLX mlx-community/granite-4.1-3b-mxfp4 for Apple silicon 4.3k downloads AWQ cyankiwi/granite-4.1-3b-AWQ-INT4 for vLLM, SGLang 14.6k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

9 rows

Benchmarks 9 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (9)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported AGI EVAL	65.16	0-shot, CoT	model author	2026-10-07
author-reported AlpacaEval · Chat and follow instructions	38.57	—	model author	2026-10-07
author-reported Arena-Hard · Chat and follow instructions	37.8	—	model author	2026-10-07
author-reported BBH · Reason through hard problems	75.83	3-shot, CoT	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	31.7	0-shot, CoT	model author	2026-10-07
author-reported IFEval · Chat and follow instructions	82.3	—	model author	2026-10-07
author-reported MMLU · Reason through hard problems	67.02	5-shot	model author	2026-10-07
author-reported MMLU-Pro · Reason through hard problems	49.83	5-shot, CoT	model author	2026-10-07
author-reported SimpleQA · Reason through hard problems	3.68	—	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

Granite-4.1-3B is a 3.40B parameter instruct model designed for general instruction following and AI assistants.^[1]^[2]

parameters human
“designed to follow general instructions”

Generated 2026-10-07 15:12 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

20 fields

Listed in LLMs

Facts

granite-4.1-3b has 3.40B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 131,072 tokens (128K). (config.json max_position_embeddings, observed 2026-10-24)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
Created 2026-04-06, last modified 2026-05-04. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	GraniteForCausalLM
Model type	granite
Dtype	bfloat16
Params by dtype	BF16: 3,402,836,480
Size band	M
MoE	no
Context length	131,072 (config.json max_position_embeddings)
Papers	arXiv:0000.00000
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · GraniteForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · GraniteForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for granite-4.1-3b: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	8.2 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	4.1 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	2.0 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Original model Family: ibm-granite/granite-4.1 107 derivatives on the Hub

finetune: 22quantized: 66adapter: 19

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF	finetune	GGUF	—	—	apache-2.0	AnkitAI	3	826.2k
ibm-granite/granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	ibm-granite	14	268.9k
unsloth/granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	unsloth	20	46.1k
ibm-granite/granite-4.1-3b-fp8	quantized	-fp8	3.66B	M	apache-2.0	ibm-granite	8	14.4k
bartowski/ibm-granite_granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	bartowski	7	14.4k
cyankiwi/granite-4.1-3b-AWQ-INT4	quantized	AWQ	3.40B	M	apache-2.0	cyankiwi	0	14.6k
lmstudio-community/granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	lmstudio-community	0	12.7k
unsloth/granite-4.1-3b	finetune	—	3.40B	M	apache-2.0	unsloth	6	11.2k
huihui-ai/Huihui-granite-4.1-3b-abliterated	quantized	GGUF	3.40B	M	apache-2.0	huihui-ai	6	11k
mlx-community/granite-4.1-3b-mxfp4	quantized	mxfp4	3.40B	M	apache-2.0	mlx-community	5	4.3k
mrutkows/granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	mrutkows	0	3.6k
mradermacher/granite-4.1-3b-i1-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	1	2.5k
AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5	finetune	—	3.40B	M	apache-2.0	AnkitAI	1	2.2k
mlx-community/granite-4.1-3b-bf16	finetune	-bf16	3.40B	M	apache-2.0	mlx-community	1	2.1k
mlx-community/granite-4.1-3b-4bit	quantized	-4bit	3.40B	M	apache-2.0	mlx-community	0	2k
mradermacher/granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	1	1.5k
ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5	adapter	—	3.40B	M	apache-2.0	ermiaazarkhalili	1	1.4k
Kezmark/Mordant-3B-Think	finetune	GGUF	3.40B	M	apache-2.0	Kezmark	1	1.4k
callmefattyy/IBM4.1-Unnoticed.Thinker.Uncensored-3B.gguf	quantized	gguf	—	—	—	callmefattyy	2	1.2k
pharos-one/pharos-judge-free	quantized	GGUF	3.40B	M	apache-2.0	pharos-one	2	1.1k
ermiaazarkhalili/Granite-4.1-3B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF	adapter	GGUF	—	—	apache-2.0	ermiaazarkhalili	0	1.2k
mlx-community/granite-4.1-3b-mxfp8	quantized	—	3.40B	M	apache-2.0	mlx-community	1	1.1k
ermiaazarkhalili/Granite-4.1-3B-Function-Calling-xLAM-Unsloth	adapter	—	3.40B	M	apache-2.0	ermiaazarkhalili	2	1.1k
ermiaazarkhalili/Granite-4.1-3B-SFT-Claude-Opus-Reasoning-Unsloth	adapter	—	3.40B	M	apache-2.0	ermiaazarkhalili	3	1k
treadon/granite-4.1-3b-Abliterated-AND-Disinhibited	finetune	—	3.40B	M	apache-2.0	treadon	4	911
Dzluck/granite-4.1-3b-GGUF	quantized	GGUF	—	—	apache-2.0	Dzluck	0	1.1k
saidutta69/granite-4.1-3b-heretic	quantized	GGUF	3.40B	M	apache-2.0	saidutta69	0	1.1k
wangzhang/granite-4.1-3b-abliterated	finetune	—	3.40B	M	apache-2.0	wangzhang	1	1k
ermiaazarkhalili/Granite-4.1-3B-Function-Calling-xLAM-Unsloth-GGUF	adapter	GGUF	—	—	apache-2.0	ermiaazarkhalili	0	991
mlx-community/granite-4.1-3b-nvfp4	quantized	nvfp4	3.40B	M	apache-2.0	mlx-community	1	933
SirSahOl/granite-4.1-3b-chat-mlx-16bit	finetune	mlx	3.40B	M	apache-2.0	SirSahOl	0	914
DevQuasar/ibm-granite.granite-4.1-3b-GGUF	quantized	GGUF	—	—	—	DevQuasar	0	905
litert-community/granite-4.1-3b	quantized	—	—	—	apache-2.0	litert-community	0	889
ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint	adapter	—	3.40B	M	apache-2.0	ermiaazarkhalili	0	888
liodon-ai/granite-4.1-3b-imatrix-GGUF	quantized	imatrix	—	—	other	liodon-ai	1	818
Kerassy/Granite-4.1-3B-Medical_Reasoning	adapter	—	3.40B	M	apache-2.0	Kerassy	1	803
ertghiu256/Granite-4.1-3b-coder	finetune	—	3.40B	M	apache-2.0	ertghiu256	1	766
yethdev/granite-4.1-3b-manumit-v1	finetune	—	3.40B	M	apache-2.0	yethdev	0	807
mlx-community/granite-4.1-3b-8bit	quantized	-8bit	3.40B	M	apache-2.0	mlx-community	0	713
phoenix-gurl/granite-4.1-3b-Q4_K_M-GGUF	quantized	GGUF	—	—	apache-2.0	phoenix-gurl	1	628

Repository changelog

14 commits

Repository changelog 14 commits

Date	Commit	SHA
2026-05-04	Granite-4.1 Model family	c065040
2026-04-28	Granite-4.1 Model family	f843067
2026-04-28	Update README.md (#1)	7a03905
2026-04-27	Granite-4.1 Model family	f3251e1
2026-04-24	Granite-4.1 Model family	6bf3697
2026-04-24	Update README.md	4ea76f1
2026-04-23	Granite-4.1 Model family	c4bc7df
2026-04-22	Update README.md	f3d419e
2026-04-22	Update README.md	6125089
2026-04-22	Update README.md	387ec8f
2026-04-21	Update README.md	8c99587
2026-04-21	Granite-4.1 Model family	f40cc4b
2026-04-08	Granite-4.1 Model family	2fdf028
2026-04-06	initial commit	3a25493

Articles and coverage

3 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official2

The IBM Granite 4.1 family of models

research.ibm.com2026-04-30officialauthorityabout the granite 4.1 familyHN 88 pts · 16 comments1,677 words

IBM released the Granite 4.1 family of models under an Apache 2.0 license, offering language models in 3B, 8B, and 30B parameter sizes alongside vision, speech, embedding, and guardian capabilities. These language models are trained on approximately 15 trillion tokens and extend context length to as much as 512K tokens, while the 97M-parameter embedding model supports retrieval across more than 200 languages. Developers can access these models on watsonx, Hugging Face, and via optimized inference runtimes including vLLM, SGLang, and llama.cpp.written, reviewed, cited

Granite Speech 4.1 2B NAR generates entire sequences at once instead of one token at a time.
The Granite 4.1 8B instruct model matches or outperforms the Granite 4.0 32B Mixture-of-Experts model.
Granite Guardian 4.1 is a direct replacement for Granite Guardian 3.3 8B and was fine-tuned on top of Granite 4.1 8B.

Granite 4.1 LLMs: How They're Built

huggingface.co2026-04-29authorityabout the granite 4.1 familyHN 1 pts · 0 comments3,171 words

Granite 4.1 is a family of dense, decoder-only LLMs (3B, 8B, and 30B) released under the Apache 2.0 license. The models are trained on approximately 15 trillion tokens using a five-phase pipeline that extends context windows up to 512K tokens. They are further refined with supervised fine-tuning on approximately 4.1 million high-quality curated samples and reinforcement learning via on-policy GRPO with DAPO loss.written, reviewed, cited

The 8B instruct model matches or surpasses the previous Granite 4.0-H-Small (32B-A9B MoE) despite using a simpler dense architecture.
The LCE phase uses an exponential learning rate schedule starting at 1e-4 and decaying to 0 for long context training.
Due to the computationally intensive nature of dynamic sampling, the team switches the DAPO loss off during training runs.

Analysis1

Granite 4.1: IBM's 8B Model Matching 32B MoE

firethering.com2026-04-30HN front pageabout the granite 4.1 familyHN 328 pts · 209 comments1,865 words

IBM released Granite 4.1, a family of open source models trained on 15 trillion tokens with a rigorous data quality pipeline. The 8B dense model outperforms the previous 32B MoE variant on benchmarks like ArenaHard and GSM8K. Development involved four sequential RL stages and resulted in 4.1 million curated samples.written, reviewed, cited

The 3B model extends context only to 128K tokens, not the 512K available to larger variants.
Phase 5 extended the context window eventually to 512K tokens for the 8B and 30B models.
Stage three was a short identity and knowledge calibration run about 40 training steps to stabilize representation.

Similar models

3 models

Similar models

Llama-3.2-3B-Instruct · 3.21B
same size band (M), same task (text-generation), more providers
Llama-3.2-3B · 3.21B
same size band (M), same task (text-generation), more providers
Qwen2.5-3B-Instruct · 3.09B
same size band (M), same task (text-generation), more providers

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:40 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:12 UTC · Last verified 2026-10-25

Follow granite-4.1-3b

Watch this model

/feeds/models/ibm-granite--granite-4.1-3b.xml(RSS 2.0) · every change to granite-4.1-3b
/feeds/families/ibm-granite--granite-4.1.xml(RSS 2.0) · every change across the granite-4.1 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.