Run and distribute models
Ollama supplies model tags, downloads, a runtime and an API. Its library tells you what is available and how to launch it.
Confidence should be
visible, not implied.
MamiLens is a decision layer, not a model host, benchmark laboratory or replacement for Ollama and Hugging Face. This page separates sourced facts from MamiLens estimates and real machine tests.
98 curated profilesReviewed 20 August 20260 claimed MamiLens lab tests
Host and share AI artifacts
Hugging Face hosts model checkpoints, datasets, demos, model cards and community repositories across many frameworks.
Choose a realistic first test
MamiLens combines a curated artifact with hardware, task and context constraints, then explains fit, uncertainty and evidence.
CLAIM POLICY
What we may and may not say.
Parameter count, published context, input types, licence, artifact size and command are checked against the linked publisher or runtime page.
A linked independent test shows that a related artifact ran on named hardware. It does not prove identical speed elsewhere.
RAM and VRAM floors are conservative planning calculations. They are not publisher guarantees or measured performance.
This label is used only after DeviceTerra runs the exact artifact, quantization, context and runtime on documented hardware. No current profile carries this claim.
ENGINE EVIDENCE
Four levels prevent false compatibility.
MamiLens now checks the engine path before workflow preferences can influence the ranking.
The selected tag or file is directly present in the engine’s official catalogue. Every current MamiLens Ollama profile meets this level.
The engine’s current official documentation lists the architecture. A separate compatible checkpoint, GGUF or MLX artifact is still required.
The engine accepts a relevant format, but MamiLens has not confirmed this exact architecture and feature path. It cannot outrank stronger evidence without a warning.
The reviewed official documentation does not establish the model-to-engine path. MamiLens excludes it from recommendations instead of guessing.
MODEL-BY-MODEL STATUS
Every current recommendation profile
“Official checked” confirms source fields. “Community linked” adds a reproducible third-party configuration. Neither means DeviceTerra personally benchmarked it.
ProfileArtifactEvidenceMamiLens testSource
Mixtral 8x22BMistral AI · MoE · 39B active
~80 GB · Q4-class Official checkedNot lab-testedQwen 3.6 35B-A3B Q4Alibaba Cloud · MoE · 3B active
~24 GB · Q4-class Official checkedNot lab-testedQwen 3.5 0.8BAlibaba Cloud · Dense
~1 GB · Ollama published artifact Official checkedNot lab-testedQwen 3.5 2BAlibaba Cloud · Dense
~2.7 GB · Ollama published artifact Official checkedNot lab-testedQwen 3.5 4BAlibaba Cloud · Dense
~3.4 GB · Ollama published artifact Official checkedNot lab-testedQwen 3.5 9BAlibaba Cloud · Dense
~6.6 GB · Ollama published artifact Official checkedNot lab-testedQwen 3.5 27BAlibaba Cloud · Dense
~17 GB · Ollama published artifact Official checkedNot lab-testedQwen 3.5 35BAlibaba Cloud · Dense
~24 GB · Ollama published artifact Official checkedNot lab-testedQwen 3.5 122BAlibaba Cloud · Dense
~81 GB · Ollama published artifact Official checkedNot lab-testedQwen 3 Coder 30B-A3.3BAlibaba Cloud · MoE · 3.3B active
~19 GB · Ollama published artifact Official checkedNot lab-testedQwen 3 Coder 480BAlibaba Cloud · Dense
~290 GB · Ollama published artifact Official checkedNot lab-testedQwen 3 30B-A3BAlibaba Cloud · MoE · 3B active
~19 GB · Q4-class Official checkedNot lab-testedQwen 3 235B-A22BAlibaba Cloud · MoE · 22B active
~142 GB · Q4-class Official checkedNot lab-testedGPT-OSS 120BOpenAI · MoE · 5.1B active
~65 GB · Native MXFP4 Official checkedNot lab-testedNemotron 3 Nano 30B-A3BNVIDIA · MoE · 3B active
~24 GB · Q4-class Official checkedNot lab-testedDeepSeek R1 1.5BDeepSeek · Dense
~1.1 GB · Q4-class Official checkedCommunity linkedNot lab-tested
DeepSeek R1 7BDeepSeek · Dense
~4.7 GB · Q4-class Official checkedCommunity linkedNot lab-tested
DeepSeek R1 8BDeepSeek · Dense
~5.2 GB · Q4-class Official checkedCommunity linkedNot lab-tested
DeepSeek R1 14BDeepSeek · Dense
~9 GB · Q4-class Official checkedCommunity linkedNot lab-tested
DeepSeek R1 32BDeepSeek · Dense
~20 GB · Q4-class Official checkedCommunity linkedNot lab-tested
DeepSeek R1 70BDeepSeek · Dense
~43 GB · Q4-class Official checkedCommunity linkedNot lab-tested
DeepSeek R1 671B MoEDeepSeek · MoE · 37B active
~404 GB · Q4-class Official checkedCommunity linkedNot lab-tested
Llama 4 Maverick 400BMeta · MoE · 17B active
~245 GB · Q4-class Official checkedNot lab-testedDevstral Small 2 24BMistral AI · Dense
~15 GB · Ollama published artifact Official checkedNot lab-testedDevstral 2 123BMistral AI · Dense
~75 GB · Ollama published artifact Official checkedNot lab-testedThe MamiLens accuracy contract
- Never convert another person’s tokens-per-second result into a promise for your computer.
- Never call a RAM estimate an official requirement.
- Never treat parameter count or one leaderboard score as proof that a model is best for every task.
- Always identify the artifact, quantization, context and runtime behind a recommendation.
- Remove or downgrade a profile when its evidence cannot be reproduced or its source changes.