Models TibiOS can run

From a compact model on a single device to a frontier model across a coordinated cluster.

TibiOS doesn't run just one AI model — it orchestrates whichever model fits the job and the hardware available, from a single edge device to a network of coordinated nodes. The table below is a sample of what's in the catalog today: the size of each model, how much memory it needs, what it's particularly good at, and roughly how capable and how fast it is.

A quick guide to what each column means:

  • Size (parameters): how many parameters the model has — very roughly, its "weight." A bigger model can hold more nuance, at the cost of needing more memory and time.
  • Memory needed: the approximate minimum memory to run the model in its most efficient compressed form.
  • Good at: the kind of task that specific model tends to do best — everyday conversation, step-by-step reasoning, writing code, or understanding images.
  • Capability level: a relative read of how capable the model is compared to the rest of the catalog — not a single universal score, since that depends heavily on the task.
  • Approximate speed: how quickly it tends to respond. Smaller models are typically near-instant on a single device; larger ones need more time, or more hardware working together.
Model Size Memory needed Good at Capability Speed
Phi-4-mini-reasoning ~3.8B params ~3 GB Step-by-step reasoning in a very compact model Solid (reasoning) Fast
Gemma 3 4B ~4.3B params ~3 GB Everyday conversation + understanding images Solid (multimodal) Fast
DeepSeek Coder 6.7B Instruct ~6.7B params ~5 GB Code generation and completion, sized for a single edge device Solid (code) Fast–Medium
DeepSeek-R1-Distill-Qwen 7B ~7.6B params ~5 GB Reasoning, math and logic — sized for a single edge device Advanced (reasoning) Fast–Medium
Mistral 7B Instruct ~7.25B params ~5 GB Everyday conversation, resource-efficient Solid Fast–Medium
Qwen3 8B ~8.2B params ~5 GB Everyday conversation, good quality/speed balance Solid Fast–Medium
Llama 3.1 8B Instruct ~8B params ~5 GB Everyday conversation, very long context (131K tokens) Solid Fast–Medium
DeepSeek-R1-Distill-Qwen 14B ~14.8B params ~9 GB Advanced reasoning, more precise than the 7B version Advanced (reasoning) Medium
Qwen3 Coder 30B-A3B ~30.5B params ~19 GB Writing and reviewing code, very long context (262K tokens) Advanced (code) Medium
DeepSeek Coder 33B Instruct ~33.3B params ~20 GB Code generation and completion, larger and more capable than the 6.7B version Advanced (code) Medium
Llama 3.3 70B Instruct ~70.6B params ~43 GB High-quality conversation, complex tasks Advanced Slower — needs a larger GPU or several nodes
GLM-4.5 ~355B params ~213 GB High-quality conversation and complex tasks, very long context (131K tokens) Frontier Slower per response — runs across a coordinated cluster
DeepSeek-V3 ~671B params ~403 GB Frontier-level capability for the most demanding tasks Frontier Slower per response — runs across a coordinated cluster
Kimi K2-Instruct ~1T params (MoE) ~600 GB Long-context, frontier-scale conversation Frontier Slower per response — runs across a coordinated cluster
Kimi K3 ~2.8T params (MoE) Not yet benchmarked — research stage Distributing giant models across many small, low-power devices instead of one big machine Research preview Not yet measured
TibiOS Sauron TBD Not yet benchmarked — in research Our own model, purpose-built for TibiOS's distributed architecture In research Not yet measured

Sizes and memory figures come from TibiOS's own model catalog. Capability and speed are relative, orientation guides, not a benchmark score — actual speed depends on the hardware running each model.