25 models. Every modality. One endpoint.
Every model below serves the same OpenAI-compatible API — pick by RAM tier
and capability, not by integration effort. Load any of these from the
console's Models tab with one click — the server downloads and brings it
online. Or pre-fetch from the shell with
fetch-models.sh (installed next to the models; also a
direct download).
The name on each card is the model id — the exact string for the API model field. The fetch-models.sh argument (its install group) can be shorter; use the card's model id in requests.
text + tools on every model vision · audio · video 4 GB to 128 GB tiers one-click console load
Qwen 3.5 family
The workhorse family — from the 0.8B speed demon to the 35B mixture-of-experts, all sharing one prompt format and one API.
q35-0.8b
Qwen 3.5 · 0.8B · needs 8 GB RAM
The fastest model in the catalog — text and tool calling only. Fits 8 GB edge hosts and doubles as a speculative-decoding draft model.
q35-2b
Qwen 3.5 · 2B · needs 16–32 GB RAM · ~1.9 GB download
Light multimodal — full vision, audio, and video understanding on a 16 GB host.
q35-4b
Qwen 3.5 · 4B · needs 32 GB RAM · ~3.4 GB download
The flagship default the installer ships — best speed/quality balance, zero-error record on the 380-prompt pack, identical results on Arm, Intel, and AMD.
q35-9b
Qwen 3.5 · 9B · needs 32 GB RAM · ~6.6 GB download
The strongest text + vision quality in the Qwen 3.5 family.
q35moe-35b
Qwen 3.5 · 35B MoE · needs 64 GB RAM
General-purpose mixture-of-experts: 35B-class answers at active-parameter speed. Audio input via the ASR sidecar.
qwen3-coder-next
Qwen 3.5 · 35B MoE · needs 64 GB RAM
Agentic coding specialist — built for long tool-calling sessions and repository-scale code work.
Qwen 3.8 family
The deep-reasoning generation: hybrid attention for long context, thinking mode, and per-request reasoning_effort control.
qwen38-27b
Qwen 3.8 · 27B · needs 64 GB RAM · ~17.2 GB download
Dense hybrid reasoner with thinking mode and reasoning_effort levels (low / medium / high) for dial-a-depth answers. Audio via the ASR sidecar; plan for a 64 GB node.
qwen38-flash-next-q2
Qwen 3.8 · 125B MoE (6B active) · needs 128 GB RAM · ~73 GB download
Frontier-scale mixture-of-experts — 125B parameters with only 6B active per token, served as UD-Q2_K_XL. Thinking model: every response carries reasoning_content. Plan for a 128 GB-class host.
Spark X2.6 family
Apache-2.0 thinking models on hosts as small as 4 GB — controllable thinking (none/low/high via reasoning_effort), strong agent/coding/long-context, 1M context, tool calling included.
spark-x26-4b
Spark X2.6 · 4B · needs 8 GB RAM · ~2.6 GB download
Apache-2.0 thinking model, same architecture as X2.5-4B with improved post-training (stronger agent, coding, long-context) — controllable thinking (none/low/high via reasoning_effort), 1M context, text and tools.
spark-x26-1.7b
Spark X2.6 · 1.7B · needs 4 GB RAM · ~1.1 GB download
Smallest controllable-thinking model (none/low/high via reasoning_effort), 1M context, text and tools on a 4 GB host. Same architecture as X2.5-1.7B, improved post-training. Apache 2.0.
Nemotron family
NVIDIA's hybrid Mamba-2 line, CPU-served — a 30B-class reasoner at active-parameter speed, a 12B vision-language model, and a sub-1B document-OCR model that turns page images into structured markdown.
nemotron-lightning-30b
Nemotron · 30B A3B MoE · needs 32 GB RAM · ~18.9 GB download
NVIDIA Nemotron-3.5 Lightning — a hybrid Mamba-2 + MoE reasoner: 30B-class quality at ~3B-active speed, thinking mode with reasoning_content in every response, text and tool calling on a 32 GB host. Served as Q4.
nemotron-vl-12b
Nemotron · 12B hybrid · needs 16 GB RAM · ~9.2 GB download
NVIDIA Nemotron-Nano-12B-v2-VL — a hybrid Mamba-2 vision-language model: describe images and answer questions about them, with a thinking mode. Image + text in, on a 16 GB host. C-RADIO vision tower.
nemotron-parse-2.0
Nemotron · <1B · needs 8 GB RAM · ~1.0 GB download
NVIDIA document-OCR — page image in, structured markdown out: text with layout classes and bounding boxes. C-RADIO vision encoder + mBART decoder, served as Q8 on an 8 GB host.
Ternary Bonsai family
Extreme-compression reasoning — 27B parameters at 1.72 bits per weight via the PQ2_0 ternary format, serving 27B-class thinking answers from a 16 GB host.
ternary-bonsai-9b
Ternary Bonsai · 9B · needs 8 GB RAM · ~3.1 GB download
Ternary-Bonsai-2 9B at ~2.2 bits/weight (PQ2_0) — 9B-class thinking answers from a 2.4 GB file on an 8 GB host. Thinking is ON by default (pass enable_thinking:false to disable). Text + images (F16 vision projector).
ternary-bonsai-4b
Ternary Bonsai · 4B · needs 4 GB RAM · ~1.8 GB download
Ternary-Bonsai-2 4B (PQ2_0) — mid-size ternary reasoning on 4-8 GB hosts. Thinking ON by default. Text + images.
ternary-bonsai-2b
Ternary Bonsai · 2B · needs 4 GB RAM · ~1.2 GB download
Ternary-Bonsai-2 2B (PQ2_0) — a thinking model in HALF A GIGABYTE. The smallest reasoning tier on the server. Thinking ON by default. Text only.
ternary-bonsai-27b
Ternary Bonsai · 27B · needs 16 GB RAM · ~7.3 GB download
Ternary-Bonsai-2 27B at 1.72 bits per weight (PQ2_0) — 27B-class reasoning in ~8 GB resident on a 16 GB host. Thinking mode with reasoning_content (pass enable_thinking:true) and reasoning_effort control. Text + images (F16 vision projector).
Strands family
Typed decisions instead of prose — POST /v1/systemone returns calibrated probabilities over your own question set in a single pass, for routing, triage, and guardrails.
strands-decider-2b
Strands · 2B · needs 8 GB RAM · ~6.2 GB download
Calibrated decision model for POST /v1/systemone — application state plus typed questions in, calibrated probabilities out (text and images). NOT a chat model: free-text generation is deliberately untrained. The Q8 variant serves in ~4 GB.
Clef family
The /v1/systemone accuracy tier — Cloudflare's 9B joint-schema decision model scores every question's options in one prefill, text and images at native resolution, calibrated probabilities out.
clef-flash-9b-q8
Clef · 9B · needs 16 GB RAM · ~10.9 GB download
Cloudflare Clef-Flash — the accuracy tier for POST /v1/systemone: state plus typed questions in, calibrated probabilities out, ALL questions scored jointly in one prefill (text and images at native resolution). A 9B post-trained Qwen3.5 trunk with a joint schema head; served as Q8. NOT a chat model.
Gemma 4 family
Native multimodal end to end — vision, audio, and video decoded in-model, no projector round-trips, on hosts as small as 8 GB.
gemma-4-E2B-it
Gemma 4 · E2B · needs 8 GB RAM · ~2.7 GB download
The smallest full-multimodal model — native text, vision, audio, and video on an 8 GB host.
gemma-4-E4B-it
Gemma 4 · E4B · needs 16 GB RAM · ~4.5 GB download
Full multimodal on a 16 GB host — 320/380 on the prompt pack, within five answers of the 4B flagship.
gemma-4-12B-it
Gemma 4 · 12B · needs 32 GB RAM · ~6.7 GB download
The media flagship: image encode in 19 ms and audio encode in ~1 ms, with native video too. The SAI_G4_W4=only profile serves it in ~7 GB RSS with instant load.
gemma-4-26B-A4B-it
Gemma 4 · 26B A4B MoE · needs 32 GB RAM · ~13.9 GB download
Highest pack quality in the catalog (329/380) at ~10 s typical answers — MoE active-parameter economics on a 32 GB host. Audio via the ASR sidecar.
Retrieval family
The search building blocks — multimodal embeddings and reranking that complete a single-binary RAG stack: embed, recall, rerank, and generate on one CPU server.
qwen3vl-embed-2b
Retrieval · 2B · needs 8 GB RAM · ~2.5 GB download
Multimodal embeddings for search and RAG — text and images share one 2048-dim vector space (OpenAI-compatible /v1/embeddings). Matryoshka dimensions 64–2048 let you shrink vectors without re-embedding; 73.2 MMEB-V2. Remote image URLs and data URLs both accepted.
qwen3vl-rerank-2b
Retrieval · 2B · needs 8 GB RAM · ~2.5 GB download
Second-stage reranker for precision — scores each document (text or image) against the query, 0–1, via /v1/rerank (Jina/Cohere-compatible). Pair it with the embedder: vector recall first, rerank the top candidates.
Pick by RAM
| Host RAM | Models that fit comfortably |
|---|---|
| 4 GB | spark-x26-1.7b |
| 8 GB | q35-0.8b, spark-x26-4b, gemma-4-E2B-it, qwen3vl-embed-2b, qwen3vl-rerank-2b, nemotron-parse-2.0, strands-decider-2b |
| 16 GB | q35-2b, gemma-4-E4B-it, nemotron-vl-12b, ternary-bonsai-27b — and gemma-4-12B-it under SAI_G4_W4=only (~7 GB RSS) |
| 32 GB | q35-4b (default), q35-9b, gemma-4-12B-it, gemma-4-26B-A4B-it, nemotron-lightning-30b |
| 64 GB | q35moe-35b, qwen3-coder-next, qwen38-27b |
| 128 GB | qwen38-flash-next-q2 |
Every model in the catalog does function calling.
score = fixed-answer classification/routing via POST /v1/score
(Qwen 3.5 and Spark models) — one prefill, no decode, deterministic.
audio* = audio input served via the ASR sidecar: install the
asr group and set asr-model — speech is transcribed
by the dedicated ASR model and handed to the text model on the same request.
The low-RAM Gemma tier (SAI_G4_W4=only) serves Gemma 12B in
~7 GB RSS with instant weight load. Sizing details in the
Getting Started guide; measured
speeds on the Performance page.