Pick a model. Copy the command. You're running.
vinOS ships Ollama. Every model below is a single ollama pull
away — no config, no auth, no keys. Runs on your GPU (or CPU).
Nothing leaves the machine. Marginal cost per prompt: your
electricity. Routines default to a local pick from this catalog and
escalate to Claude only when they have to — see the router below.
Updated daily.
A GitHub Action
runs at 06:00 UTC, diffs this catalog against
ollama.com/library,
and opens an issue when new models appear so a human can curate
them in. Model space moves fast; the page tries to keep up.
The 80/20 router
Every vinOS routine picks a model per run. Set route = "auto"
in the routine's TOML and the runtime tries a local Ollama
model first — for summarization, extraction, drafting,
classification, small-scale reasoning. Roughly 80% of routine
work never leaves your machine. The other 20% — deep reasoning,
big context, code that must actually be right — escalates to a
premium API (Claude Sonnet/Opus). Escalation happens when the local
model returns low confidence, when the task declares itself as
reasoning-heavy, or when the local model runs out of context.
The math on a typical founder day: 500 routine
runs/day (github-review three times × ~30 PRs, inbox-triage hourly
× ~15 emails, day-brief once, evening-shutdown once, ad-hoc
vinos-ai calls). 20% escalate to Claude
at roughly $0.03 per premium call average
(Sonnet input+output, ~2k in / ~500 out). That's
$90/mo in Claude
spend. If you routed everything to Claude —
$450/mo.
Saved: ~$360/mo
per user, or roughly a MacBook every eighteen months.
Configure the router
Drop route = "auto" into the routine's [agent]
block and declare a local_model, a premium_model,
and an escalation policy. That's it.
[agent]
route = "auto" # anthropic | ollama | auto
local_model = "qwen2.5:7b" # what runs the 80% case
premium_model = "claude-sonnet-4-6" # what runs the escalated 20%
[agent.escalation]
on_low_confidence = true # local returns “unsure” or confidence < threshold
on_reasoning_task = true # prompt tagged as reasoning-heavy
on_context_overflow = true # local hit its context limit
max_escalations_per_run = 3 # cap runaway loops
Set route = "ollama" to force local-only (zero API spend).
Set route = "anthropic" when the routine's whole reason
for being is high-stakes judgment. Full field reference:
routine spec.
Just tell me what to install.
Pick the row that matches your machine. Two commands and you're set for daily use — general chat + coding assistant.
4 GB RAM
ollama pull qwen2.5:1.5b8 GB RAM
ollama pull llama3.2:3b && ollama pull qwen2.5-coder:3b16 GB RAM · recommended
ollama pull llama3.2:8b && ollama pull qwen2.5-coder:7b32 GB + GPU
ollama pull qwen2.5:32b && ollama pull qwen2.5-coder:32b64+ GB · workstation
ollama pull llama3.3:70b100+ GB · fat server
ollama pull deepseek-v3Then: Super+A opens chat with your default model. Or vinos-ai chat in any terminal.
How much are you saving?
Local marginal cost per prompt: $0 (your electricity — call it a rounding error). The "cost vs cloud" column on each table below compares against the nearest frontier equivalent — API rates for reference:
Claude Opus 4.7
$15/M in · $75/M out — deep reasoning
Claude Sonnet 4.6
$3/M in · $15/M out — general purpose
Claude Haiku 4.5
$1/M in · $5/M out — fast + cheap
GPT-4 class
roughly comparable to Opus
text-embedding-3
$0.02/M — for context vs Ollama's $0
One heavy user's typical week: ~$150/mo saved when 80% of prompts route local.
General chat & instruct
Day-to-day Super+A chat, drafting, summarizing, extraction, structured tasks.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1Kimi K2 · Moonshot AI huggingface · ollama.com new | 1T params · 32B active (MoE) 128k ctx · Jul 2025 | 89.5MMLU · 1T MoE65.8 SWE-bench Verified | Moonshot AI's frontier open MoE — tops open-model coding + agentic benches. Trained as a native tool-user; ships with strong function-calling. 32B active per token → runs on a beefy workstation despite the 1T total. Under-the-radar in the West, dominant in Chinese AI-agent stacks. Modified MIT (Kimi K2) | 32B active → ~20 GB (Q4) · 1T → ~500 GB full | $0 vs ~$0.20 / agent step (Opus) | ollama pull kimi-k2 |
| 2Hunyuan Large · Tencent ollama.com | 7B · 389B (MoE, ~52B active) 256k ctx · Nov 2024 | 88.4MMLU · 389B MoE | Tencent's open MoE flagship. Beats Llama 3.3 70B and Claude Sonnet on multiple benches. Massive but only 52B active per token. Tencent Hunyuan license | 52B active → 32 GB · 389B → 220 GB | $0 vs ~$0.20 / prompt (Opus) | ollama pull hunyuan |
| 2Llama 3.3 · Meta ollama.com | 70B 128k ctx · Dec 2024 | 86.0MMLU · 70B68.9 MMLU Pro | Top-tier open reasoning. GPT-4-class on many benches. Llama 3.3 license | 70B → 40 GB | $0 vs ~$1.65 / 100-pg summary (Opus) | ollama pull llama3.3 |
| 3Llama 4 Maverick · Meta ollama.com | 400B params · 17B active (MoE) 1M ctx · Apr 2026 | 85.5MMLU · MoE80.5 MMLU Pro | Meta's flagship native-multimodal MoE. Datacenter class. Llama 4 license | 400B → ~200 GB | $0 vs GPT-4-class API rates | ollama pull llama4:maverick |
| 4Phi 4 · Microsoft ollama.com | 14B 16k ctx · Dec 2024 | 84.8MMLU | Punches above weight class. Reasoning-focused training. MIT | 14B → 10 GB | $0 vs ~$0.20 / prompt (Opus) | ollama pull phi4 |
| 5Mistral Large · Mistral ollama.com | 123B 128k ctx · Jul 2024 | 84.0MMLU | Mistral's flagship. For serious hardware. Mistral Research (non-commercial) | 123B → 68 GB | $0 vs ~$0.20 / prompt (Opus) | ollama pull mistral-large |
| 6Qwen 3 · Alibaba ollama.com | 0.6B · 1.7B · 4B · 8B · 14B · 30B · 32B · 235B 40k–256k ctx · May 2025 | 83.3MMLU · 32B | Latest Qwen. Hybrid thinking/non-thinking modes. Widest size range on Ollama. Apache 2.0 | 8B → 6 GB · 32B → 20 GB | $0 vs ~$0.05 / prompt (Sonnet) | ollama pull qwen3 |
| 7Qwen 2.5 · Alibaba ollama.com | 0.5B · 1.5B · 3B · 7B · 14B · 32B · 72B 128k ctx · Sep 2024 | 83.3MMLU · 32B | Sharp on structured output (JSON, tables). Best multilingual. Widest size range. Apache 2.0 | 7B → 5 GB · 32B → 20 GB | $0 vs ~$1.50 / 500-email extract (Opus) | ollama pull qwen2.5:7b |
| 8GLM-4 · Zhipu AI ollama.com | 9B · 32B · 100B 128k ctx · Jun 2024 | 81.9MMLU · 32B | Zhipu AI's flagship. Sharp on Chinese + English. Widely deployed in Asia; competitive with Claude Sonnet on multiple benches. GLM-4 open license | 32B → 20 GB | $0 vs ~$0.05 / prompt (Sonnet) | ollama pull glm4 |
| 9Llama 4 Scout · Meta ollama.com | 109B params · 17B active (MoE) 10M ctx · Apr 2026 | 79.6MMLU · MoE | Massive-context multimodal at 17B active — accessible for local rigs. Llama 4 license | 109B → ~50 GB (Q4) | $0 vs long-context Sonnet | ollama pull llama4:scout |
| 10Gemma 3 · Google ollama.com | 1B · 4B · 12B · 27B 128k ctx · Mar 2025 | 78.6MMLU · 27B | Newer Gemma. Multimodal-capable at higher sizes. Gemma terms | 12B → 8 GB | $0 vs ~$0.05 / prompt (Sonnet) | ollama pull gemma3 |
| 11Mixtral 8x22B · Mistral ollama.com | 141B params · ~39B active 64k ctx · Apr 2024 | 77.3MMLU | Larger Mixtral MoE. Excellent general purpose for beefy hardware. Apache 2.0 | 141B → 80 GB | $0 vs ~$0.20 / prompt (Opus) | ollama pull mixtral:8x22b |
| 12Yi · 01.AI ollama.com | 6B · 9B · 34B 200k ctx · Jan 2024 | 76.3MMLU · 34B | Chinese-first with strong bilingual EN. Apache 2.0 | 9B → 7 GB | $0 vs ~$0.03 / prompt (Sonnet) | ollama pull yi |
| 13Command R+ · Cohere ollama.com | 104B 128k ctx · Aug 2024 | 75.7MMLU | Cohere flagship. Enterprise-tuned. CC-BY-NC 4.0 | 104B → 60 GB | $0 vs ~$0.25 / RAG call (Opus) | ollama pull command-r-plus |
| 14Gemma 2 · Google ollama.com | 2B · 9B · 27B 8k ctx · Jun 2024 | 75.2MMLU · 27B | Strong reasoning benches per parameter. Google Research lineage. Gemma terms | 27B → 16 GB | $0 vs ~$0.03 / prompt (Sonnet) | ollama pull gemma2:27b |
| 15DBRX · Databricks ollama.com | 132B params · 36B active (MoE) 32k ctx · Mar 2024 | 73.7MMLU | Databricks fine-grained MoE. Enterprise-grade + Apache-safe. DBRX Open Model License | 132B → 74 GB | $0 vs enterprise API rates | ollama pull dbrx |
| 16Falcon 3 · TII ollama.com | 1B · 3B · 7B · 10B 32k ctx · Dec 2024 | 73.1MMLU · 10B | UAE's Falcon. Sovereign-friendly. 10B leads sub-13B knowledge tasks. TII Falcon 2.0 | 10B → 7 GB | $0 vs cloud | ollama pull falcon3 |
| 17Mistral Small · Mistral ollama.com | 22B 32k ctx · Sep 2024 | 72.9MMLU | Mid-tier from Mistral. Balances speed + quality. Mistral Research (non-commercial) | 22B → 14 GB | $0 vs ~$0.05 / prompt (Sonnet) | ollama pull mistral-small |
| 18InternLM 2.5 ollama.com | 1.8B · 7B · 20B 1M ctx · Jul 2024 | 72.8MMLU · 7B | Shanghai AI Lab. Long context (1M tokens at 7B). Apache 2.0 | 7B → 5 GB | $0 vs ~$1+ / 1M-context call (Sonnet) | ollama pull internlm2 |
| 19Gemma 2 9B · Google ollama.com | 9B 8k ctx · Jun 2024 | 71.3MMLU · 9B | Sweet-spot Gemma 2 for consumer boxes. Gemma terms | 9B → 7 GB | $0 vs ~$0.03 / prompt (Sonnet) | ollama pull gemma2:9b |
| 20Llama 3.1 · Meta ollama.com | 8B · 70B · 405B 128k ctx · Jul 2024 | 69.4MMLU · 8B | The 8B daily driver. Still the strongest small default for most users. Llama 3.1 license | 8B → 6 GB | $0 vs ~$0.20 / refactor (Opus) | ollama pull llama3.1:8b |
| 21Command R · Cohere ollama.com | 35B · 08-2024 128k ctx · Aug 2024 | 68.2MMLU | Tuned for RAG + tool-use. Long context. CC-BY-NC 4.0 | 35B → 22 GB | $0 vs ~$0.15 / RAG call (Sonnet) | ollama pull command-r |
| 22Mistral Nemo · Mistral ollama.com | 12B 128k ctx · Jul 2024 | 68.0MMLU | 128k context window. Long-document work. Apache 2.0 | 12B → 8 GB | $0 vs ~$0.80 / long-context prompt (Sonnet) | ollama pull mistral-nemo |
| 23Nous Hermes 3 ollama.com | 3B · 8B · 70B · 405B 128k ctx · Aug 2024 | 68.0MMLU · 8B | Nous Research fine-tunes of Llama. Function-calling + reasoning tuned. Llama 3.x terms | 8B → 6 GB | $0 vs ~$0.10 / tool call (Sonnet) | ollama pull hermes3 |
| 24Solar · Upstage ollama.com | 10.7B 4k ctx · Feb 2024 | 65.5MMLU | Upstage's depth-upscaled Llama. Strong benches for its size. Apache 2.0 | 10.7B → 8 GB | $0 vs ~$0.04 / prompt (Sonnet) | ollama pull solar |
| 25OpenChat ollama.com | 7B 8k ctx · Jan 2024 | 64.5MMLU | C-RLFT fine-tune. Community-loved conversational quality. Apache 2.0 | 7B → 5 GB | $0 vs ~$0.02 / prompt (Haiku) | ollama pull openchat |
| 26Mistral 7B · Mistral ollama.com | 7B 32k ctx · May 2024 | 62.5MMLU | Classic Mistral. Still a strong lightweight baseline. Apache 2.0 | 7B → 5 GB | $0 vs ~$0.03 / prompt (Sonnet) | ollama pull mistral |
| 27Aya Expanse · Cohere ollama.com | 8B · 32B 128k ctx · Oct 2024 | 62.2MMLU · 8B | 23 languages. Best truly-multilingual open model. CC-BY-NC 4.0 | 8B → 6 GB | $0 vs ~$0.05 / translation (Sonnet) | ollama pull aya-expanse |
| 28Zephyr · HuggingFace ollama.com | 7B · 141B 32k ctx · Nov 2023 | 61.1MMLU | HuggingFace's Mistral fine-tune. Strong for chat. MIT | 7B → 5 GB | $0 vs ~$0.02 / prompt (Haiku) | ollama pull zephyr |
Coding assistants
Wire into vinos-ai code or your editor. Fill-in-the-middle, refactor, explain.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1Qwen 2.5 Coder · Alibaba ollama.com | 0.5B · 1.5B · 3B · 7B · 14B · 32B 128k ctx · Nov 2024 | 92.7HumanEval · 32B90.2 MBPP | Top open coding model. 32B beats Claude 3.5 Sonnet on HumanEval. Recommended default. Apache 2.0 | 7B → 5 GB · 32B → 20 GB | $0 vs ~$0.20 / typical refactor (Opus) | ollama pull qwen2.5-coder:7b |
| 2DeepSeek Coder V2 ollama.com | 16B (MoE) · 236B 128k ctx · Jun 2024 | 90.2HumanEval · 236B76.2 MBPP | MoE — activates ~2.4B/token (Lite) or ~21B (236B). Great on larger codebases. DeepSeek license | 16B → 10 GB | $0 vs ~$1+ / complex refactor (Opus) | ollama pull deepseek-coder-v2 |
| 3Codestral · Mistral ollama.com | 22B 32k ctx · May 2024 | 81.1HumanEval78.2 MBPP | Solid completion + fill-in-the-middle. 32k context. Mistral Research (non-commercial) | 22B → 14 GB | $0 vs ~$0.15 / completion (Sonnet) | ollama pull codestral |
| 4DeepSeek Coder ollama.com | 1.3B · 6.7B · 33B 16k ctx · Nov 2023 | 79.3HumanEval · 33B | Predecessor to V2. Still solid; different license terms. DeepSeek license | 6.7B → 5 GB | $0 vs ~$0.15 / refactor (Sonnet) | ollama pull deepseek-coder |
| 5Magicoder ollama.com | 6.7B · 7B 4k ctx · Dec 2023 | 76.8HumanEval | OSS-Instruct fine-tune. Very community-loved. CC-BY-4.0 | 7B → 5 GB | $0 vs ~$0.10 / prompt (Sonnet) | ollama pull magicoder |
| 6Codestral Mamba · Mistral ollama.com | 7B 256k ctx · Jul 2024 | 75.0HumanEval | Mamba architecture — infinite context, fast inference. Apache 2.0 | 7B → 5 GB | $0 vs ~$0.15 / long-file completion (Sonnet) | ollama pull codestral-mamba |
| 7WizardCoder ollama.com | Python variants: 7B · 13B · 33B 16k ctx · Aug 2023 | 73.2HumanEval · 34B | Evol-Instruct fine-tune. Historical footprint but still useful. Llama 2 terms | 13B → 9 GB | $0 vs cloud | ollama pull wizardcoder |
| 8StarCoder2 · BigCode ollama.com | 3B · 7B · 15B 16k ctx · Feb 2024 | 72.6HumanEval · 15B-Instruct | Trained on The Stack v2. Strong on obscure languages. BigCode OpenRAIL-M | 15B → 10 GB | $0 vs ~$0.10 / language port (Sonnet) | ollama pull starcoder2:15b |
| 9Granite Code · IBM ollama.com | 3B · 8B · 20B · 34B 128k ctx · May 2024 | 53.0HumanEval · 8B | IBM's enterprise-grade code model. Apache 2.0 = commercial-safe. Apache 2.0 | 8B → 6 GB | $0 vs enterprise API rates | ollama pull granite-code:8b |
| 10CodeLlama · Meta ollama.com | 7B · 13B · 34B · 70B 16k ctx · Aug 2023 | 48.8HumanEval · 34B | Older but battle-tested. Python-specialized variants exist. Llama 2 license | 13B → 9 GB | $0 vs ~$0.20 / refactor (Opus) | ollama pull codellama:13b |
Reasoning specialists
Chain-of-thought native. Slower per token, deeper answers. Great for math, planning, multi-step problems.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1DeepSeek R1 ollama.com | 1.5B · 7B · 8B · 14B · 32B · 70B · 671B 128k ctx · Jan 2025 | 97.3MATH-500 · 671B79.8 AIME 2024 | DeepSeek's reasoning flagship + distills. Best math/code chain-of-thought open model. MIT | 7B → 5 GB · 671B → 400+ GB | $0 vs ~$5+ / reasoning task (Opus) | ollama pull deepseek-r1:7b |
| 2DeepSeek R1 Distill Qwen 32B ollama.com | 32B 128k ctx · Jan 2025 | 94.3MATH-500 · 32B distill72.6 AIME 2024 | The best reasoning-per-GB ratio. Fits on a single 24GB card. MIT | 32B → 20 GB | $0 vs ~$5+ / reasoning task (Opus) | ollama pull deepseek-r1:32b |
| 3QwQ · Alibaba ollama.com | 32B 32k ctx · Nov 2024 | 90.6MATH-50050.0 AIME | Qwen's reasoning-tuned variant. Chain-of-thought native. Apache 2.0 | 32B → 20 GB | $0 vs ~$5+ / deep task (Opus) | ollama pull qwq |
| 4DeepSeek V3.1 ollama.com | 671B (MoE) 160k ctx · Aug 2025 | 90.4MATH-500 | V3 refresh — hybrid thinking/non-thinking mode. Fat-server only. DeepSeek license | 671B → ~400 GB | $0 vs GPT-4-class API rates | ollama pull deepseek-v3.1 |
| 5DeepSeek V3 ollama.com | 685B (MoE) 128k ctx · Dec 2024 | 90.2MATH-500 | DeepSeek's MoE. Extremely capable but needs a fat server. DeepSeek license | 685B → ~400 GB (MoE) | $0 vs GPT-4-class API rates | ollama pull deepseek-v3 |
| 6Reflection Llama ollama.com | 70B 8k ctx · Sep 2024 | 89.9GSM8K | Self-correcting reflection tokens. Numbers were disputed at launch — treat with caution. Llama 3.1 terms | 70B → 40 GB | $0 vs ~$5+ / uncertain task (Opus) | ollama pull reflection |
| 7Marco-o1 · Alibaba ollama.com | 7B 32k ctx · Nov 2024 | 82.5MGSM | Alibaba's OpenAI-o1-style reasoning fine-tune. Compact + strong on math. Apache 2.0 | 7B → 5 GB | $0 vs ~$3+ / math problem (Opus) | ollama pull marco-o1 |
Small / edge (≤ 4B parameters)
Old laptops, low-RAM boxes, background daemons, edge devices. Instant streaming, modest capability.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1Phi 3.5 mini | 3.8B 128k ctx · Aug 2024 | 69.0MMLU | Quality-per-parameter leader. Solid at code + reasoning. MIT | 3.8B → 3 GB | $0 vs ~$0.005 / prompt (Haiku) | ollama pull phi3.5 |
| 2Llama 3.2 · Meta | 1B · 3B 128k ctx · Sep 2024 | 63.4MMLU · 3B | Instant streaming for autocomplete, short-form chat. Llama 3.2 license | 3B → 3 GB | $0 vs ~$0.005 / prompt (Haiku) | ollama pull llama3.2:3b |
| 3MiniCPM 3 | 4B 32k ctx · Sep 2024 | 62.6MMLU | Runs on mobile. Chinese lightweight lineage, strong at its size. Apache 2.0 | 4B → 3 GB | $0 vs ~$0.005 / prompt (Haiku) | ollama pull minicpm3 |
| 4Qwen 2.5 1.5B | 0.5B · 1.5B 128k ctx · Sep 2024 | 60.9MMLU · 1.5B | Runs on a Raspberry Pi 5 or a 2015 laptop. Apache 2.0 | 1.5B → 2 GB | $0 vs ~$0.005 / prompt (Haiku) | ollama pull qwen2.5:1.5b |
| 5Falcon 3 (small) | 1B · 3B 32k ctx · Dec 2024 | 56.0MMLU · 3B | TII's small Falcons. Sovereign-friendly. TII Falcon 2.0 | 3B → 3 GB | $0 vs cloud | ollama pull falcon3:3b |
| 6Gemma 2 2B | 2B 8k ctx · Jun 2024 | 52.2MMLU | Efficient, strong per-parameter reasoning. Gemma terms | 2B → 2 GB | $0 vs ~$0.005 / prompt (Haiku) | ollama pull gemma2:2b |
| 7SmolLM 2 · HuggingFace ollama.com | 135M · 360M · 1.7B 8k ctx · Oct 2024 | 49.8MMLU · 1.7B | HuggingFace's tiny series. Sub-billion sizes for embedded work. Apache 2.0 | 1.7B → 2 GB | $0 vs edge API rates | ollama pull smollm2 |
| 8Stable LM 2 ollama.com | 1.6B · 12B 4k ctx · Feb 2024 | 47.1MMLU · 1.6B | Stability AI's LM series. 1.6B for edge, 12B for mid-tier. Stability commercial terms | 1.6B → 2 GB | $0 vs ~$0.005 / prompt (Haiku) | ollama pull stablelm2 |
| 9TinyLlama ollama.com | 1.1B 2k ctx · Jan 2024 | 25.3MMLU | Pre-trained on 3T tokens. Tiny but well-trained. Apache 2.0 | 1.1B → 2 GB | $0 vs edge API rates | ollama pull tinyllama |
Vision & multimodal
Pass an image (screenshot, chart, document scan) with the prompt.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1Qwen 2.5 VL · Alibaba ollama.com | 3B · 7B · 32B · 72B 32k ctx · Jan 2025 | 70.2MMMU · 72B74.8 MathVista | Sharp on documents + charts + tables + agent screens. 72B ~= GPT-4o. Apache 2.0 | 7B → 6 GB · 72B → 45 GB | $0 vs ~$0.02 / image (Sonnet Vision) | ollama pull qwen2.5-vl |
| 2Llama 3.2 Vision · Meta ollama.com | 11B · 90B 128k ctx · Sep 2024 | 60.3MMMU · 90B57.3 MathVista | Meta's multimodal. Charts, screenshots, docs. Llama 3.2 license | 11B → 8 GB · 90B → 55 GB | $0 vs ~$0.02 / image (Sonnet Vision) | ollama pull llama3.2-vision |
| 3MiniCPM-V ollama.com | 8B 32k ctx · Jul 2024 | 49.8MMMU val | Efficient multimodal for mobile / edge. Strong OCR. Apache 2.0 | 8B → 6 GB | $0 vs ~$0.02 / OCR page | ollama pull minicpm-v |
| 4LLaVA ollama.com | 7B · 13B · 34B 4k ctx · Jan 2024 | 44.7MMMU · 34B | Classic community vision assistant. Well-documented. Apache 2.0 | 7B → 5 GB | $0 vs ~$0.02 / image (Sonnet Vision) | ollama pull llava |
| 5LLaVA-Llama3 ollama.com | 8B 8k ctx · May 2024 | 41.7MMMU | LLaVA architecture on Llama 3. Faster, sharper than v1 LLaVA. Llama 3 license | 8B → 6 GB | $0 vs ~$0.02 / image (Sonnet Vision) | ollama pull llava-llama3 |
| 6BakLLaVA ollama.com | 7B 4k ctx · Oct 2023 | 39.0MMMU | Mistral-based multimodal. Alternative lineage to LLaVA. Apache 2.0 | 7B → 5 GB | $0 vs ~$0.02 / image | ollama pull bakllava |
| 7Moondream ollama.com | 1.8B 2k ctx · Jan 2024 | —OCR-focused | Tiny vision model, runs on phones. Real-time edge tasks. Apache 2.0 | 1.8B → 2 GB | $0 vs ~$0.01 / image (Haiku Vision) | ollama pull moondream |
Embeddings
Text → vectors for search, RAG, semantic dedup. Fast, small, no chat.
| Model | Dims | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1BGE-M3 · BAAI ollama.com | 1024 8k ctx · Jan 2024 | 66.9MTEB mixed | Multilingual, multi-granularity. Dense + sparse + multi-vector in one model. MIT | 2 GB | $0 vs multi-vector API rates | ollama pull bge-m3 |
| 2mxbai-embed-large · Mixed Bread ollama.com | 1024 512 ctx · Mar 2024 | 64.7MTEB avg | High-quality general-purpose embedder. Strong bench numbers. Apache 2.0 | 1 GB | $0 vs $0.13 / M tokens (Voyage) | ollama pull mxbai-embed-large |
| 3BGE Large EN · BAAI ollama.com | 1024 512 ctx · Oct 2023 | 64.2MTEB avg | English-focused. High MTEB scores. MIT | 1 GB | $0 vs $0.02 / M tokens | ollama pull bge-large |
| 4Granite Embedding · IBM ollama.com | 384 · 768 512 ctx · Dec 2024 | 62.5MTEB avg | IBM's enterprise-grade embeddings. Apache = commercial-safe. Apache 2.0 | 1 GB | $0 vs enterprise embedding rates | ollama pull granite-embedding |
| 5Nomic Embed Text ollama.com | 768 8k ctx · Feb 2024 | 62.3MTEB avg | SOTA open embeddings for general text. Fast, tokenizer-friendly. Apache 2.0 | 1 GB | $0 vs $0.02 / M tokens (text-embedding-3) | ollama pull nomic-embed-text |
| 6Paraphrase Multilingual ollama.com | 768 128 ctx · Jun 2022 | 58.4MIRACL | Sentence-Transformers lineage. 50+ languages. Apache 2.0 | 1 GB | $0 vs $0.10 / M multilingual (Cohere) | ollama pull paraphrase-multilingual |
| 7Snowflake Arctic Embed ollama.com | 1024 512 ctx · Apr 2024 | 55.9MTEB retrieval | Multi-domain retrieval. Strong on business docs + code. Apache 2.0 | 1 GB | $0 vs $0.02 / M tokens (text-embedding-3) | ollama pull snowflake-arctic-embed |
Function-calling & agent tuning
Tuned specifically for tool-use, JSON output, and structured agent workflows.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1Firefunction v2 ollama.com | 70B 8k ctx · Jul 2024 | 85.7BFCL v2 | Fireworks AI's function-calling tune of Llama 3. GPT-4-class tool use. Llama 3 license | 70B → 40 GB | $0 vs ~$0.10 / tool call (Sonnet) | ollama pull firefunction-v2 |
| 2Hermes 2 Pro · Nous ollama.com | 7B · 8B · 70B 8k ctx · Mar 2024 | 75.4BFCL v2 | Tool-use tuned Llama. Native JSON mode. Llama terms | 8B → 6 GB | $0 vs ~$0.05 / call (Sonnet) | ollama pull hermes3 |
| 3Nexus Raven ollama.com | 13B 16k ctx · Jan 2024 | 74.8BFCL v2 | Function-calling on CodeLlama base. Compact. Llama 2 license | 13B → 9 GB | $0 vs ~$0.08 / tool call | ollama pull nexusraven |
| 4Command R (again) | 35B 128k ctx · Aug 2024 | 68.2MMLU | Cohere's RAG + tool-use flagship at accessible size. See General table. CC-BY-NC 4.0 | 35B → 22 GB | $0 vs ~$0.15 / RAG call | ollama pull command-r |
Mixture-of-Experts (MoE)
Sparse — activates only a fraction of parameters per token. High capability, moderate active compute.
| Model | Config | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1DeepSeek V3 (again) | 685B params · ~37B active 128k ctx · Dec 2024 | 88.5MMLU | The current OSS MoE flagship. Only for fat servers. DeepSeek license | ~400 GB | $0 vs GPT-4-class rates | ollama pull deepseek-v3 |
| 2Llama 4 Maverick (again) | 400B params · 17B active 1M ctx · Apr 2026 | 85.5MMLU · MoE | Meta's flagship native-multimodal MoE. Datacenter class. Llama 4 license | ~200 GB | $0 vs GPT-4-class rates | ollama pull llama4:maverick |
| 3Llama 4 Scout (again) | 109B params · 17B active 10M ctx · Apr 2026 | 79.6MMLU · MoE | Massive-context multimodal at 17B active — fits a fat consumer rig. Llama 4 license | ~50 GB (Q4) | $0 vs long-context Sonnet | ollama pull llama4:scout |
| 4DeepSeek Coder V2 (MoE) (again) | 236B params · ~21B active 128k ctx · Jun 2024 | 79.2MMLU · 236B90.2 HumanEval | MoE coding model. Very fast for its capability. See Coding table for HumanEval. DeepSeek license | 16B → 10 GB | $0 vs ~$1 / complex refactor | ollama pull deepseek-coder-v2 |
| 5Mixtral 8x22B (Mistral) | 141B params · ~39B active 64k ctx · Apr 2024 | 77.3MMLU | Larger MoE. Excellent general purpose for beefy hardware. Apache 2.0 | 141B → 80 GB | $0 vs ~$0.20 / prompt (Opus) | ollama pull mixtral:8x22b |
| 6DBRX (Databricks) (again) | 132B params · 36B active 32k ctx · Mar 2024 | 73.7MMLU | Databricks fine-grained MoE. Enterprise-focused. DBRX Open Model License | 132B → 74 GB | $0 vs enterprise API rates | ollama pull dbrx |
| 7Mixtral 8x7B (Mistral) | 47B params · ~13B active 32k ctx · Dec 2023 | 70.6MMLU | Original MoE. Still strong. GPT-3.5-class at fraction of compute. Apache 2.0 | 47B → 26 GB | $0 vs ~$0.05 / prompt (Sonnet) | ollama pull mixtral |
Uncensored / research
Alignment removed for research, red-teaming, or specific unrestricted workflows. Not for production without your own guardrails.
| Model | Sizes | Score | Best for | RAM (Q4) | Cost vs cloud | Get it |
|---|---|---|---|---|---|---|
| 1Nous Hermes 2 Mixtral ollama.com | 8x7B 32k ctx · Jan 2024 | 71.8MMLU · Mixtral base | Nous Research fine-tune on Mixtral. Very steerable. Apache 2.0 | 26 GB | $0 no cloud equivalent | ollama pull nous-hermes2-mixtral |
| 2Dolphin Mixtral ollama.com | 8x7B · 8x22B 32k ctx · Feb 2024 | 70.6MMLU · Mixtral base | Uncensored fine-tune of Mixtral. High capability + no refusal. Apache 2.0 | 8x7B → 26 GB | $0 no cloud equivalent | ollama pull dolphin-mixtral |
| 3Dolphin 3 ollama.com | 8B 8k ctx · Nov 2024 | 66.5MMLU · Llama 3 base | Eric Hartford's uncensored Llama 3 fine-tune. Community classic. Llama 3 license | 8B → 6 GB | $0 no cloud equivalent | ollama pull dolphin3 |
| 4Dolphin Llama 3 ollama.com | 8B · 70B 8k ctx · May 2024 | 66.5MMLU · Llama 3 base | Uncensored Llama 3 fine-tune. Alignment stripped. Llama 3 license | 8B → 6 GB | $0 no cloud equivalent | ollama pull dolphin-llama3 |
| 5WizardLM 2 Uncensored ollama.com | 7B · 8x22B 32k ctx · Apr 2024 | 61.2MMLU · Mistral base | Evol-Instruct fine-tune with alignment removed. Llama terms | 7B → 5 GB | $0 no cloud equivalent | ollama pull wizardlm-uncensored |
How to use these on vinOS.
Once the ai bundle is installed (vinos-install ai), Ollama is already running. Then:
ollama pull <model>Download the weights (one-time, per model).
ollama run <model>Interactive chat in the terminal.
vinos-ai chatvinOS wrapper. Uses your default model. Bound to Super+A.
vinos-ai modelsList everything you've pulled.
vinos-ai statusOllama up? Which model is default? What's using RAM?
ollama rm <model>Delete weights to reclaim disk.
ollama lsList installed models + sizes on disk.
This list will be automated in v1.2.
Right now this is a curated snapshot maintained by hand. A future vinos-models command will populate this table live from the Ollama registry plus a vinOS-blessed metadata layer (RAM estimates, cost anchors, workload tags, license badges) — and, more importantly, will recommend the right model for your specific machine and workflow. RAG pipeline? Code assistant? Long-context research? Edge deployment? The current picker is the raw material; the v1.2 personalization layer is what turns it into an answer. Model missing? File a PR against site/content/models.md.