Models › Fine-tune a base

Fine-tune a base

For fine-tuning: permissive original models under 40B, by what the community already built on them.

Fastest first token114msDeepSeek-V4.1-FlashTogether AI via huggingface.co, 2026-10-26
Cheapest with tool calls$0.010per 1M inQwen3-Embedding-8BDeepInfra via huggingface.co, 2026-10-26
Most fine-tuned 2025 base2.6kfine-tunesQwen3-4B-Instruct-2507Hugging Face base-model tags, 2026-10-26

Original (not derived) models under 40B with a permissive licence, released since 2025, sorted by how many fine-tunes the community has built on them.

The floor

Every view leaves out test fixtures, models whose size didn't parse, and closed models (unless you ask for them). $0 endpoints show as a free tier and never count as cheapest; classifiers and guards never enter a speed sort. How the rank works.

114 models match, first 60 shown. Data observed 2026-10-26 07:36 UTC, Hub snapshot 2026-10-26. Hover a figure for its source and date.

More columns

Fine-tune a base: 114 models
#
1	Qwen3-4B-Instruct-2507Qwen	4.02B	2.6k	8.9k	256K	Apache 2.0
2	Qwen2.5-VL-7B-InstructQwen	8.29B	1.3k	1.8k	125K	Apache 2.0
3	Qwen3-0.6B-BaseQwen	596M	735	924	32K	Apache 2.0
4	Qwen3.8-27BQwen	27.8B	723	2.6k	256K	Apache 2.0
5	DeepSeek-R1-Distill-Qwen-1.5Bdeepseek-ai	1.78B	705	1.3k	128K	MIT
6	Qwen3-VL-8B-InstructQwen	8.77B	677	1k	256K	Apache 2.0
7	Qwen3-8B-BaseQwen	8.19B	637	850	32K	Apache 2.0
8	Qwen3.5-9B-BaseQwen	9.65B	633	771	256K	Apache 2.0
9	Qwen3-32BQwen	32.8B	593	1.3k	40K	Apache 2.0
10	gpt-oss-20bopenai	20.9B	567	1.3k	128K	Apache 2.0
11	Qwen3-4B-BaseQwen	4.02B	513	701	32K	Apache 2.0
12	Qwen3-VL-4B-InstructQwen	4.44B	504	847	256K	Apache 2.0
13	Qwen3-1.7B-BaseQwen	1.72B	477	611	32K	Apache 2.0
14	functiongemma-270m-itgoogle	268M	471	572	32K	Gemma Terms of Use
15	gemma-3-1b-ptgoogle	1000M	427	524	32K	Gemma Terms of Use
16	Qwen3.6-27BQwen	27.8B	416	1.8k	256K	Apache 2.0
17	Qwen3.6-35B-A3BQwen	36.0B 3B active	396	1.7k	256K	Apache 2.0
18	gemma-3-4b-ptgoogle	4.30B	348	394	—	Gemma Terms of Use
19	Qwen3.5-27BQwen	27.8B	338	709	256K	Apache 2.0
20	DeepSeek-R1-Distill-Qwen-7Bdeepseek-ai	7.62B	314	743	128K	MIT
21	embeddinggemma-300mgoogle	303M	293	626	2K	Gemma Terms of Use
22	Qwen3-VL-2B-InstructQwen	2.13B	284	498	256K	Apache 2.0
23	Qwen3-4B-Thinking-2507Qwen	4.02B	272	511	256K	Apache 2.0
24	Qwen3.5-4B-BaseQwen	4.66B	240	394	256K	Apache 2.0
25	layaconvaiinnovations	421M	202	313	—	Apache 2.0
26	DeepSeek-R1-Distill-Llama-8Bdeepseek-ai	8.03B	173	671	128K	MIT
27	gemma-3-270mgoogle	268M	165	279	32K	Gemma Terms of Use
28	gemma-4-E2Bgoogle	5.12B	150	244	128K	Apache 2.0
29	Qwen3.5-0.8B-BaseQwen	873M	149	264	256K	Apache 2.0
30	gemma-4-E4Bgoogle	8.00B	143	199	128K	Apache 2.0
31	Phi-4-mini-instructmicrosoft	3.84B	136	542	128K	MIT
32	embeddinggemma-2google	744M	134	257	—	Apache 2.0
33	DeepSeek-OCRdeepseek-ai	3.34B	131	172	8K	MIT
34	Qwen3-30B-A3B-Instruct-2507Qwen	30.5B 3B active	123	404	256K	Apache 2.0
35	Qwen3.5-2B-BaseQwen	2.27B	120	197	256K	Apache 2.0
36	gemma-3-12b-ptgoogle	12.2B	116	214	—	Gemma Terms of Use
37	Qwen3-14B-BaseQwen	14.8B	108	182	32K	Apache 2.0
38	SmolLM3-3B-BaseHuggingFaceTB	3.08B	104	157	—	Apache 2.0
39	Qwen3-Coder-30B-A3B-InstructQwen	30.5B 3B active	98	347	256K	Apache 2.0
40	gemma-4-31Bgoogle	31.3B	98	153	256K	Apache 2.0
41	DeepSeek-R1-Distill-Qwen-14Bdeepseek-ai	14.8B	97	415	128K	MIT
42	DeepSeek-R1-Distill-Qwen-32Bdeepseek-ai	32.8B	91	500	128K	MIT
43	gemma-4-12Bgoogle	12.0B	89	158	256K	Apache 2.0
44	gemma-4-26B-A4Bgoogle	25.8B 4B active	87	129	256K	Apache 2.0
45	laya-multilingualconvaiinnovations	322M	85	130	—	Apache 2.0
46	Qwen3.5-35B-A3B-BaseQwen	36.0B 3B active	85	109	—	Apache 2.0
47	GLM-4.7-Flashzai-org	31.2B	81	208	198K	MIT
48	Qwen2.5-Omni-7BQwen	10.7B	81	168	—	Apache 2.0 (apache-2.0)
49	Qwen3-30B-A3B-BaseQwen	30.5B 3B active	81	144	—	Apache 2.0
50	Qwen3-VL-8B-ThinkingQwen	8.77B	79	149	256K	Apache 2.0
51	MiniCPM5-2Bopenbmb	2.52B	79	227	128K	Apache 2.0
52	Qwen3-VL-32B-InstructQwen	33.4B	74	144	256K	Apache 2.0
53	Qwen2.5-VL-32B-InstructQwen	33.5B	73	115	125K	Apache 2.0
54	DeepSeek-R1-0528-Qwen3-8Bdeepseek-ai	8.19B	68	207	128K	MIT
55	MiniCPM5-1Bopenbmb	1.08B	60	224	128K	Apache 2.0
56	Muse-Glimmer-30Bmeta-models	29.8B	55	263	128K	Apache 2.0
57	OLMo-2-0425-1Ballenai	1.48B	47	73	4K	Apache 2.0
58	Qwen3-30B-A3B-Thinking-2507Qwen	30.5B 3B active	45	141	256K	Apache 2.0
59	Qwen3-Omni-30B-A3B-InstructQwen	35.3B 3B active	41	99	—	Apache 2.0 (apache-2.0)
60	Ornith-1.5-9Bornith-ai	9.65B	40	158	256K	MIT

Evidence marks, left to right: coding, tools, reasoning, long docs, vision, chat. measured on a third-party board, author-reported, structural, none yet. Each mark links to its job from the model page. INT4 is an estimate of the weights alone (method).