Models · generated from the repo

Every model goinfer runs, and how sure we are about each one.

Each family is checked against HuggingFace's own implementation, and the result is printed next to it. A few checkpoints are vetted end to end: pinned, sized, fitted and timed. Everything else says what hasn't been shown yet.

model families

checked against the released model

checked against a small test model only

with no check of its own

vetted checkpoints

Fit and speed for

Vetted checkpoints: the supported tier

These are the checkpoints the project supports. Each is pinned by checksum and comes from a family checked against its released model, the strongest of the checks, with a fit verdict and whatever speed was measured. Other checkpoints of a family run through the same code with less shown for them. If you don't know where to start, start here.

Qwen2.5-Coder 0.5B

q4_k_m · 0.49 GB

fitsfitsfits

Code completion and small edits

Against the released model

171.2 tok/s 1.18× Ollama367.9 tok/s 1.40× Ollama52 tok/s 0.91× Ollamano speeds for a machine we haven't measured

Qwen2.5-Coder 1.5B

q4_k_m · 1.1 GB

fitsfitsfits

Qwen2 and Qwen2.5, including the Coder models most of goinfer's numbers are measured on.

Against the released model

89.3 tok/s 1.06× Ollama270.7 tok/s 1.47× Ollama24.3 tok/s level with Ollamano speeds for a machine we haven't measured

Phi-3 mini 4k

q4 · 2.4 GB

fits on the CPUfits on the CPUfits

General instruction following at a size that still loads in seconds

Against the released model

not measured on this machine112.6 tok/s 0.90× Ollamanot measured on this machineno speeds for a machine we haven't measured

Granite 4.0-H Tiny

q8_0 · 7.4 GB

fits on the CPUfits on the CPUfits

A Mamba-2/attention hybrid MoE, if you want to exercise that path

Against the released model

not measured on this machinenot measured on this machinenot measured on this machineno speeds for a machine we haven't measured

gpt-oss 20B

mxfp4 · 12.1 GB

tight pages experts from diskfits experts streamed to the cardfits

The 20-35B-class MoE this project actually validates and measures

Against the released model

not measured on this machinenot measured on this machinenot measured on this machineno speeds for a machine we haven't measured

Gemma 4 26B-A4B

q4_0 · 14.4 GB

tight pages experts from diskfits experts streamed to the cardfits

The 20-35B-class MoE this project has the most measurements on

Against the released model

not measured on this machine39.3 tok/s 1.76× Ollama, graded ambiguousnot measured on this machineno speeds for a machine we haven't measured

All families

Sorted by how strong the check is, strongest first. There's no "popular" sort: goinfer doesn't count downloads.

For

Checked

Hide experimental

Needs goinfer can download it (GGUF) Runs on a GPU Mixture of experts Has a vetted checkpoint

Showing 40 of 40 familiesWhat the checks mean

Embeddings

Separate from the families above: they are encoders, not chat models.

goinfer serve --embed-model serves /v1/embeddings, on its own or beside a chat model. Two encoders work today. Qwen3-Embedding-0.6B (a GGUF file) matches the sentence-transformers reference to cosine 1.0000000 on five test inputs, checked 2026-07-20. CodeRankEmbed (a HuggingFace directory) is served and timed against Ollama's nomic-embed-text, but no comparison with a reference is recorded here. Details in the server docs.

What the checks mean

Every family gets one of these. They are listed strongest first.

Against the released model

The family's output is compared with HuggingFace's own implementation on a real checkpoint. Two numbers are printed: how often the next token is the same, and the closest the raw scores get at their worst position (cosine, where 1.00000 is identical).

Against a small test model

No released model was small enough to run on the hardware available, so the family is checked against a small test model built from the same wiring. A weaker claim, and it is labelled that way.

Shares another family's check

The family runs another family's code, so that family's numbers stand in for it. It has none of its own.

Not recorded

No check is on file for this family yet.

More in what "parity-gated" means. A family registered as experimental can change, or go, before v1.0.

Coming from Ollama?

Of the 60 most-pulled models in Ollama's library on 2026-10-02, goinfer runs the families behind 93.3% of the pulls as Ollama ships them, and another 1.0% as text only.

Pull counts are cumulative, so older models rank higher than their current use; this is a guide to what loads, not a measure of what people run.

#	Ollama tag	goinfer	Family	Note
1	`llama3.1`	Runs	Llama
2	`deepseek-r1`	Runs	Qwen2 / Qwen2.5, Qwen3, Llama, DeepSeek-V3	distills plus the 671B
3	`nomic-embed-text`	Runs	Embeddings
4	`llama3.2`	Runs	Llama
5	`qwen2.5`	Runs	Qwen2 / Qwen2.5
6	`gemma3`	Runs	Gemma 3
7	`qwen3`	Runs	Qwen3
8	`gemma2`	Runs	Gemma 2
9	`mistral`	Runs	Mistral
10	`gemma4`	Runs	Gemma 4	images yes; audio input not yet
11	`llama3`	Runs	Llama
12	`qwen2.5-coder`	Runs	Qwen2 / Qwen2.5
13	`qwen3.5`	Runs	Qwen3.8, Qwen3.5-MoE	images from safetensors, dense sizes only; Ollama's GGUF and the MoE sizes run as text
14	`phi3`	Runs	Phi-3 / Phi-4
15	`mxbai-embed-large`	Runs	Embeddings
16	`llava`	Not supported yet
17	`gpt-oss`	Runs	gpt-oss
18	`qwen3-coder`	Runs	Qwen3-MoE
19	`gemma`	Runs	Gemma	Gemma 1
20	`qwen`	Not supported yet		Qwen 1

The other 40

#	Ollama tag	goinfer	Family	Note
21	`glm-ocr`	Runs	GLM-OCR	images from the safetensors checkpoint only (no GGUF); the vision tower runs on the CPU; tool calls are not rendered; parity tier experimental
22	`phi4`	Runs	Phi-3 / Phi-4	Phi-4 uses the phi3 architecture
23	`llama2`	Runs	Llama
24	`bge-m3`	Runs	Embeddings
25	`qwen3.6`	Runs	Qwen3.8, Qwen3.5-MoE	images from safetensors, dense sizes only; Ollama's GGUF and the MoE sizes run as text
26	`codellama`	Runs	Llama
27	`qwen3-vl`	Runs as text only	Qwen3-VL	text side experimental; safetensors only
28	`qwen2`	Runs	Qwen2 / Qwen2.5
29	`tinyllama`	Runs	Llama
30	`mistral-nemo`	Runs	Mistral
31	`minicpm-v`	Not supported yet
32	`llama3.2-vision`	Not supported yet		cross-attention (mllama)
33	`qwen2.5vl`	Runs	Qwen2.5-VL	safetensors only
34	`deepseek-coder`	Runs	Llama
35	`llama3.3`	Runs	Llama
36	`qwen3-embedding`	Runs	Qwen3
37	`dolphin3`	Runs	Llama
38	`smollm2`	Runs	Llama
39	`deepseek-v3`	Runs	DeepSeek-V3
40	`olmo2`	Not supported yet		goinfer has olmo3, not olmo2
41	`all-minilm`	Runs	Embeddings
42	`codegemma`	Runs	Gemma	the Gemma 1 architecture
43	`deepseek-coder-v2`	Runs	DeepSeek-V2
44	`mistral-small`	Runs	Mistral
45	`snowflake-arctic-embed`	Runs	Embeddings
46	`orca-mini`	Runs	Llama
47	`granite3.1-moe`	Not supported yet		granitemoe; goinfer has granite and granitemoehybrid
48	`qwen3.8`	Runs	Qwen3.8, Qwen3.5-MoE	images from safetensors, dense sizes only; Ollama's GGUF and the MoE sizes run as text
49	`starcoder2`	Not supported yet
50	`nemotron-3-super`	Not checked yet	Nemotron-H	family matches; the Super MoE variant is unchecked
51	`mixtral`	Runs	Mixtral	parity tier experimental
52	`llama2-uncensored`	Runs	Llama
53	`falcon3`	Runs	Llama
54	`translategemma`	Runs	Gemma 3
55	`mistral-small3.2`	Runs as text only	Ministral 3	vision tower ignored; safetensors only
56	`minimax-m2.7`	Not supported yet		minimax_m2
57	`llava-llama3`	Not supported yet
58	`embeddinggemma`	Runs	Gemma 3
59	`qwq`	Runs	Qwen2 / Qwen2.5
60	`gemma3n`	Not supported yet

Ollama's library page, read 2026-10-02. The statuses and the figures come from the snapshot.