Catalogue

every model we serve, and what it really costs

One OpenAI-compatible endpoint, open weights only, priced per million tokens with no minimum and no subscription. Each model reaches several classes of GPU through different quantizations, so a 6 GB card and a 24 GB card both have work to do — and hosts keep 80% of what their hardware earns.

what your GPU can serve

Your card	Models	For example
6 GB	3	Gemma 4 E4B Instruct, Qwen2.5 1.5B Instruct, Qwen2.5 0.5B Instruct
8 GB	3	Gemma 4 E4B Instruct, Qwen2.5 1.5B Instruct, Qwen2.5 0.5B Instruct
10 GB	5	Qwen3.8 27B Instruct, Gemma 4 12B Instruct, Gemma 4 E4B Instruct
12 GB	7	Qwen3.8 27B Instruct, GLM-4.7-Flash, Qwen3 Coder 30B-A3B
16 GB	9	Qwen3.8 27B Instruct, Qwen3.6 35B-A3B, GLM-4.7-Flash
24 GB	10	Qwen3.8 27B Instruct, Qwen3.6 35B-A3B, GLM-4.7-Flash

Minimum VRAM per build. Install mahout and run mahout models — it tells you exactly what your own card can serve.

Qwen3.8 27B Instruct

qwen3.8-27b

chatcodereasoninglong-contextmultilingualtool-usevision

27B dense 256K Apache-2.0 from 9 GB €0.25 in · €0.75 out per million tokens

2026 flagship open model for a single card: strong reasoning and code, 262K context, and no EU-sovereign listing anywhere else.

9 builds — one per VRAM class

quantization	download	needs	context	notes
`UD-Q5_K_M`	18.4 GB	23 GB VRAM	32K	reference quality
`UD-Q4_K_M`	15.3 GB	20 GB VRAM	32K	reference quality
`UD-IQ4_XS`	13.3 GB	18 GB VRAM	32K	balanced
`UD-Q3_K_XL`	12.2 GB	16 GB VRAM	32K	balanced
`UD-IQ3_S`	11.2 GB	15 GB VRAM	32K	balanced
`UD-IQ3_XXS`	10.2 GB	14 GB VRAM	32K	compact — fits smaller cards
`UD-Q2_K_XL`	9.2 GB	12 GB VRAM	32K	compact — fits smaller cards
`UD-IQ2_S`	7.8 GB	11 GB VRAM	16K	compact — fits smaller cards
`UD-IQ2_XXS`	6.8 GB	9 GB VRAM	16K	compact — fits smaller cards

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Qwen3.6 35B-A3B

qwen3.6-35b-a3b

chatcodereasoninglong-contextmultilingualtool-usevision

35B MoE, 3B active 256K Apache-2.0 from 14 GB €0.18 in · €0.55 out per million tokens

Fast enough to delegate to: 3B active parameters decode at small-model speed while the full 35B answers, which is what makes it the one to hand sub-agent work to. Drafts ahead, sees images, 262K context.

8 builds — one per VRAM class

quantization	download	needs	context	notes
`UD-IQ4_NL`	17.3 GB	22 GB VRAM	32K	reference quality
`UD-IQ4_XS`	17 GB	21 GB VRAM	32K	reference quality
`UD-Q3_K_XL`	16 GB	20 GB VRAM	32K	balanced
`UD-IQ3_S`	14.3 GB	19 GB VRAM	32K	balanced
`UD-IQ3_XXS`	13.1 GB	18 GB VRAM	32K	balanced
`UD-Q2_K_XL`	11.7 GB	16 GB VRAM	32K	compact — fits smaller cards
`UD-IQ2_XXS`	11 GB	15 GB VRAM	32K	compact — fits smaller cards
`UD-IQ1_M`	10.6 GB	14 GB VRAM	16K	compact — fits smaller cards

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

GLM-4.7-Flash

glm-4.7-flash

chatcodereasoninglong-contextmultilingualtool-use

31B MoE 198K MIT from 11 GB €0.16 in · €0.5 out per million tokens

MIT licensed end to end — the most permissive terms in the catalogue, and the answer for buyers whose legal team reads the licence. Strong general reasoning from a second vendor.

9 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_1`	17.7 GB	23 GB VRAM	32K	reference quality
`UD-Q4_K_XL`	16.3 GB	21 GB VRAM	32K	reference quality
`IQ4_XS`	15.2 GB	20 GB VRAM	32K	balanced
`UD-Q3_K_XL`	12.8 GB	17 GB VRAM	32K	balanced
`UD-IQ3_XXS`	12 GB	16 GB VRAM	32K	balanced
`UD-Q2_K_XL`	11.1 GB	15 GB VRAM	32K	compact — fits smaller cards
`UD-IQ2_XXS`	9.8 GB	14 GB VRAM	32K	compact — fits smaller cards
`UD-IQ1_S`	8.6 GB	12 GB VRAM	16K	compact — fits smaller cards
`UD-TQ1_0`	7.8 GB	11 GB VRAM	16K	compact — fits smaller cards

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Qwen3 Coder 30B-A3B

qwen3-coder-30b-a3b

chatcodereasoninglong-contexttool-use

30B MoE, 3B active 256K Apache-2.0 from 11 GB €0.2 in · €0.6 out per million tokens

Built for code and agentic tool use: repository-scale context at 262K, and 3B active parameters so it answers at conversational speed.

9 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_1`	17.9 GB	23 GB VRAM	32K	reference quality
`UD-Q4_K_XL`	16.5 GB	21 GB VRAM	32K	reference quality
`IQ4_XS`	15.3 GB	20 GB VRAM	32K	balanced
`UD-Q3_K_XL`	12.9 GB	17 GB VRAM	32K	balanced
`UD-IQ3_XXS`	12 GB	16 GB VRAM	32K	balanced
`UD-Q2_K_XL`	11 GB	15 GB VRAM	32K	compact — fits smaller cards
`UD-IQ2_XXS`	9.6 GB	14 GB VRAM	32K	compact — fits smaller cards
`UD-IQ1_M`	9 GB	13 GB VRAM	16K	compact — fits smaller cards
`UD-TQ1_0`	7.5 GB	11 GB VRAM	16K	compact — fits smaller cards

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Gemma 4 26B-A4B Instruct

gemma-4-26b-a4b

chatreasoningmultilinguallong-context

26B MoE, 4B active 128K Gemma Terms of Use from 17 GB €0.15 in · €0.45 out per million tokens

Mixture-of-experts: 26B of knowledge at the speed of a 4B model. Quantization-aware weights published by Google, so the 4-bit build loses far less than a post-hoc quantization.

1 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_0 (QAT)`	13.4 GB	17 GB VRAM	32K	reference quality

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

gpt-oss 20B

gpt-oss-20b

chatcodetool-usereasoning

20B MoE 128K Apache-2.0 from 14 GB €0.12 in · €0.4 out per million tokens

Open-weight 20B tuned for tool use and structured output. Natively MXFP4, so 4-bit costs it almost nothing.

1 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_K_M`	10.8 GB	14 GB VRAM	32K	reference quality

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Gemma 4 12B Instruct

gemma-4-12b

chatreasoningmultilingual

12B dense 128K Gemma Terms of Use from 10 GB €0.08 in · €0.25 out per million tokens

The 8 GB workhorse: a capable general assistant on cards most gamers already own, with Google's quantization-aware 4-bit weights.

1 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_0 (QAT)`	6.5 GB	10 GB VRAM	32K	reference quality

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Gemma 4 E4B Instruct

gemma-4-e4b

chatmultilingual

8B MatFormer, 4B effective 128K Gemma Terms of Use from 6 GB €0.05 in · €0.15 out per million tokens

Built for small cards and laptops: 6 GB of VRAM is enough to earn.

1 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_0 (QAT)`	4.8 GB	6 GB VRAM	16K	reference quality

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Qwen2.5 1.5B Instruct

qwen2.5-1.5b-instruct

chat

1.5B dense 32K Apache-2.0 from 3 GB €0.05 in · €0.15 out per million tokens

Tiny, fast, runs anywhere — the enrollment smoke test and the model behind starter work for a freshly linked machine.

1 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_K_M`	1 GB	3 GB VRAM	32K	reference quality

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

Qwen2.5 0.5B Instruct

qwen2.5-0.5b-instruct

chat

0.5B dense 32K Apache-2.0 from 2 GB €0.02 in · €0.06 out per million tokens

Smallest catalog entry: CPU-only fallback and multi-model swap tests.

1 builds — one per VRAM class

quantization	download	needs	context	notes
`Q4_K_M`	0.5 GB	2 GB VRAM	32K	reference quality

Every build is pinned to an exact upstream revision and verified against its SHA-256 before it runs — on your machine if you host, and on the machine serving you if you build.

use them, or serve them

Point your SDK's base URL at https://app.elephantpool.ai/api/gateway/v1 and pass the model id — or connect a GPU and get paid for every token it produces.

Serve with your GPU Call the API