Supported models

jitllm supports a fixed list of architectures, each one checked against a reference implementation rather than assumed from matching tensor names. Any GGUF or Hugging Face checkpoint of a listed architecture converts, not only the ones in the catalog.

Architectures

Dense transformers

Architecture	Models
llama	Llama 2, 3, 3.1, 3.2, 3.3; Mistral 7B; SmolLM2; TinyLlama; the text models of SmolVLM and LLaVA; ERNIE 4.5 0.3B (ernie4_5)
mistral3	Mistral Small 3.1/3.2, Magistral, Devstral, Ministral 3
qwen2, qwen3	Qwen2, Qwen2.5, Qwen3; Qwen3-Embedding
gemma, gemma2, gemma3, gemma3n, gemma4	Gemma 1, 2, 3, 3n and 4; EmbeddingGemma
phi2, phi3	Phi-2; Phi-3, Phi-3.5, Phi-4
glm4	GLM-4-0414 (9B, 32B), GLM-Z1, and the text models of GLM-4.1V and GLM-4.6V-Flash
granite	IBM Granite 3.x and 4.0
command-r, cohere2	Command-R, Aya, Command-R7B, Command-A
smollm3, arcee, apertus	SmolLM3, Arcee AFM-4.5B, Apertus
olmo2, exaone4, seed_oss	OLMo 2 and OLMo 3, EXAONE 4.0, Seed-OSS
hunyuan-dense	Tencent Hunyuan (0.5B to 7B)
starcoder, starcoder2	StarCoder, StarCoderBase, StarCoder2
stablelm, falcon, nemotron	StableLM 2, Falcon, Nemotron

Mixtures of experts

Architecture	Models
qwen3moe, qwen2moe	Qwen3 MoE; Qwen1.5-MoE, Qwen2-57B-A14B
olmoe, dbrx, phimoe	OLMoE, DBRX, Phi-3.5-MoE; Mixtral 8x7B through the llama graph
gpt-oss	gpt-oss 20B and 120B
llama4	Llama 4 Scout (text)
granitemoe	IBM Granite mixtures
glm4moe	GLM-4.5, GLM-4.6, GLM-4.5-Air, and the text models of GLM-4.5V and GLM-4.6V
ernie4_5-moe, hunyuan-moe	ERNIE 4.5 (21B-A3B, 300B-A47B), Hunyuan mixtures
minimax-m2, minimax-m3	MiniMax-M2, M2.1, M2.5 and M3 (text)
bailingmoe2, dots1	Ling and Ring 2.0, dots.llm1
deepseek2, deepseek32, deepseek4	DeepSeek-V2, V3, R1, V3.2 and V4; DeepSeek-Coder-V2; Kimi K2; Moonlight; GLM-4.7-Flash; Kimi-VL (latent attention)

Hybrids and state-space models

Architecture	Models
qwen3next, qwen35, qwen35moe	Qwen3-Next, Qwen3.5, Qwen3.6: gated linear attention beside attention
Kimi-Linear, kimi-k3	Kimi-Linear (from safetensors), Kimi-K3 (text): Kimi Delta Attention beside latent attention
mamba, mamba2	Mamba-1 and FalconMamba; Mamba-2 and Mamba-Codestral
jamba, falcon-h1	Jamba; Falcon-H1
granitehybrid	IBM Granite 4.0-H
nemotron_h, nemotron_h_moe	Nemotron-H, Nemotron Nano 2, Nemotron 3 Nano
lfm2, lfm2moe	Liquid LFM2 and LFM2-8B-A1B

Vision and embeddings

Architecture	Models
qwen2vl, qwen3vl, qwen3vlmoe	Qwen2-VL, Qwen2.5-VL, Qwen3-VL (2B to 235B-A22B)
bert, nomic-bert	sentence-embedding encoders (all-MiniLM, mxbai-embed, nomic-embed)

Vision towers convert beside their language model into one container, and their blocks page and place like any other block.

Projector	Families
idefics3	SmolVLM, Idefics3
mlp	LLaVA-style models (a CLIP or SigLIP tower and a two-layer MLP)
qwen2vl_merger, qwen2.5vl_merger, qwen3vl_merger	Qwen2-VL, Qwen2.5-VL, Qwen3-VL
glm4v	GLM-4.1V, GLM-4.5V, GLM-4.6V
gemma3, gemma3nv, gemma4v	Gemma 3, Gemma 3n (MobileNet-V5), Gemma 4
internvl	InternVL 2.5 and 3
resampler	MiniCPM-V 2.6, 4.0, 4.5
kimivl	Kimi-VL (MoonViT)
janus_pro, pixtral, phi4, llama4	Janus-Pro; Pixtral and Mistral 3; Phi-4 reasoning vision; Llama 4

From Hugging Face safetensors, these model classes convert directly: Llama, Mixtral, Qwen2, Qwen2Moe, Qwen3, Qwen3Moe, Olmoe, Olmo2, Olmo3, Gemma, Phi3, PhiMoE, DeepseekV2, DeepseekV3, Glm4Moe, Glm4MoeLite, KimiLinear, Ernie4_5, Ernie4_5_Moe, HunYuanDenseV1, HunYuanMoEV1, Dots1, Apertus, SmolLM3, Arcee, SeedOss, Exaone4 and Ministral3. Every other architecture converts from GGUF.

An architecture that is not on the list is refused at conversion with its name, not converted into something that runs and produces wrong text.

Weight formats

Format	Notes
F32, F16, BF16	unquantized; -q8 turns a safetensors model’s matrices into Q8_0
Q8_0, Q4_0, Q5_0, Q5_1	legacy block formats
Q3_K, Q4_K, Q5_K, Q6_K	K-quants, as in every Q4_K_M or Q5_K_M file
MXFP4	gpt-oss’s expert weights

Conversion keeps the source’s quantization and changes only its layout, so a .jlm is about the size of the GGUF it came from.

The catalog

jitllm convert NAME out.jlm fetches these by name, pinned to a revision checked against this engine. jitllm library prints the same list.

Name	Parameters	Architecture	Weights	Download
smollm2-360m-instruct	360M	llama	Q8_0	386 MB
llama-3.2-1b-instruct	1B	llama	Q4_K_M	808 MB
smollm2-1.7b-instruct	1.7B	llama	Q4_K_M	1.1 GB
llama-3.2-3b-instruct	3B	llama	Q4_K_M	2.0 GB
mistral-7b-instruct-v0.3	7B	llama	Q4_K_M	4.4 GB
llama-3.1-8b-instruct	8B	llama	Q4_K_M	4.9 GB
llama-3.3-70b-instruct	70B	llama	Q4_K_M	42.5 GB
qwen2.5-1.5b-instruct	1.5B	qwen2	Q4_K_M	1.1 GB
qwen2.5-7b-instruct	7B	qwen2	Q4_K_M	4.7 GB
qwen2-vl-2b-instruct	2B	qwen2vl	Q4_K_M	1.7 GB	images
qwen3-0.6b	0.6B	qwen3	Q8_0	639 MB
qwen3-1.7b	1.7B	qwen3	Q8_0	1.8 GB
qwen3-4b	4B	qwen3	Q4_K_M	2.5 GB
qwen3-8b	8B	qwen3	Q4_K_M	5.0 GB
qwen3-14b	14B	qwen3	Q4_K_M	9.0 GB
qwen3-30b-a3b	30B-A3B	qwen3moe	Q4_K_M	18.6 GB
qwen3-next-80b-a3b-instruct	80B-A3B	qwen3next	Q4_K_M	48.5 GB
qwen3.5-2b	2B	qwen35	Q4_K_M	1.3 GB
qwen3.5-4b	4B	qwen35	Q4_K_M	2.7 GB
qwen3.6-35b-a3b	35B-A3B	qwen35moe	Q4_K_M	22.1 GB
phi-2	2.7B	phi2	Q4_K_M	1.8 GB
starcoderbase-1b	1B	starcoder	Q4_K_M	793 MB
aya-expanse-8b	8B	command-r	Q4_K_M	5.1 GB
command-r7b	7B	cohere2	Q4_K_M	5.1 GB
starcoder2-3b	3B	starcoder2	Q4_K_M	1.9 GB
falcon-7b-instruct	7B	falcon	Q4_K_M	5.0 GB
qwen1.5-moe-a2.7b-chat	14.3B-A2.7B	qwen2moe	Q4_K_M	9.5 GB
ernie-4.5-21b-a3b	21B-A3B	ernie4_5-moe	Q4_K_M	13.3 GB
ernie-4.5-0.3b	0.3B	ernie4_5	Q4_K_M	241 MB
hunyuan-0.5b-instruct	0.5B	hunyuan-dense	Q4_K_M	355 MB
ling-mini-2.0	16B-A1.4B	bailingmoe2	Q4_K_M	9.9 GB
phi-mini-moe-instruct	7.6B-A2.4B	phimoe	Q4_K_M	5.0 GB
apertus-8b-instruct	8B	apertus	Q4_K_M	5.1 GB
granite-4.0-h-350m	350M	granitehybrid	Q8_0	366 MB
nemotron-nano-9b-v2	9B	nemotron_h	Q4_K_M	6.5 GB
nemotron-3-nano-30b-a3b	30B-A3B	nemotron_h_moe	Q4_K_M	24.7 GB
falcon-h1-0.5b-instruct	0.5B	falcon-h1	Q4_K_M	315 MB
lfm2-350m	350M	lfm2	Q4_K_M	229 MB
lfm2-8b-a1b	8B-A1B	lfm2moe	Q4_K_M	5.0 GB
mamba-codestral-7b	7B	mamba2	Q4_K_M	4.1 GB
falcon-mamba-7b	7B	mamba	Q4_K_M	4.2 GB
jamba-reasoning-3b	3B	jamba	F16	6.4 GB
nemotron-mini-4b	4B	nemotron	Q4_K_M	2.7 GB
smollm3-3b	3B	smollm3	Q4_K_M	1.9 GB
afm-4.5b	4.5B	arcee	Q4_K_M	2.9 GB
seed-oss-36b	36B	seed_oss	Q4_K_M	21.8 GB
olmo-2-1b	1B	olmo2	Q4_K_M	936 MB
exaone-4.0-1.2b	1.2B	exaone4	Q4_K_M	812 MB
ministral-3-3b	3B	mistral3	Q4_K_M	2.1 GB
stablelm-2-1.6b	1.6B	stablelm	Q4_K_M	1.0 GB
granite-3.3-2b	2B	granite	Q4_K_M	1.5 GB
granite-4.0-micro	3B	granite	Q4_K_M	2.1 GB
granite-3.1-1b-a400m	1B-A400M	granitemoe	Q4_K_M	822 MB
deepseek-v2-lite	16B-A2.4B	deepseek2	Q4_K_M	10.4 GB
deepseek-coder-v2-lite-instruct	16B-A2.4B	deepseek2	Q4_K_M	10.4 GB
moonlight-16b-a3b-instruct	16B-A3B	deepseek2	Q4_K_M	10.5 GB
olmoe-1b-7b-instruct	7B-A1B	olmoe	Q4_K_M	4.2 GB
gemma-2-2b-it	2B	gemma2	Q4_K_M	1.7 GB
gemma-2-9b-it	9B	gemma2	Q4_K_M	5.8 GB
gemma-3-1b-it	1B	gemma3	Q4_K_M	806 MB
gemma-3-4b-it	4B	gemma3	Q4_K_M	2.5 GB
gemma-3-12b-it	12B	gemma3	Q4_K_M	7.3 GB
gemma-3n-E2B-it	5B (E2B)	gemma3n	Q4_K_M	3.0 GB
phi-3.5-mini-instruct	3.8B	phi3	Q4_K_M	2.4 GB
phi-4-mini-instruct	3.8B	phi3	Q4_K_M	2.5 GB
phi-4	14B	phi3	Q4_K_M	9.1 GB
llama-4-scout-17b-16e-instruct	109B-A17B	llama4	Q3_K_S	46.7 GB
gpt-oss-20b	21B-A3.6B	gpt-oss	MXFP4	12.1 GB
gpt-oss-120b	117B-A5.1B	gpt-oss	MXFP4	63.4 GB
all-minilm-l6-v2	22M	bert	F16	46 MB	embeddings
nomic-embed-text-v1.5	137M	nomic-bert	F16	274 MB	embeddings
mxbai-embed-large-v1	335M	bert	F16	670 MB	embeddings
qwen3-embedding-0.6b	0.6B	qwen3	F16	1.2 GB	embeddings
smolvlm-256m-instruct	256M	llama	Q8_0	279 MB	images
smolvlm-500m-instruct	500M	llama	Q8_0	546 MB	images
llava-phi-3-mini	3.8B	llama	int4	2.9 GB	images
qwen3-vl-2b-instruct	2B	qwen3vl	Q4_K_M	1.6 GB	images
qwen3-vl-30b-a3b-instruct	30B-A3B	qwen3vlmoe	Q4_K_M	19.3 GB	images
glm-4.6v-flash	9B	glm4	Q4_K_M	7.1 GB	images
glm-4-9b-0414	9B	glm4	Q4_K_M	6.2 GB
kimi-vl-a3b-thinking-2506	16B-A3B	deepseek2	Q4_K_M	11.2 GB	images

The largest (Llama 3.3 70B, Qwen3-Next 80B, gpt-oss 120B) are larger than the memory of an ordinary desktop. They run anyway, paged from disk, at a speed set by the drive.