Is this a ranking of the best AI models?

No. Maviter lists a finite set of exact, revision-pinned model records so people can compare evidence and local-fit constraints without treating popularity as quality.

Does Maviter host or redistribute model files?

No. Maviter links to the publisher repository and does not copy model weights into its application or GitHub repository.

Are the memory figures guaranteed requirements?

No. Formula-backed rows are planning estimates at 4-bit weights and an 8,192-token context. Runtime, quantisation, cache implementation and hardware can materially change real usage.

Maviter Model Directory / revision 2026-09-01.1

Compare the exact model—not the family name.

Ten publisher records, pinned to exact revisions. Filter by the work, input type, licence, context and realistic hardware class before you download anything or choose an inference route.

Explore exact models →Check your RAM and VRAM

MODEL DECISION / 01Evidence before download

Exact IDContextMemoryReal test

1.0BGoogle3.84BMicrosoft4.30BGoogle4.02BQwen

Circle size indicates parameter class only. It is not a quality score.

10exact model releases

5publisher groups

8formula-backed memory rows

0files mirrored by Maviter

Know what you are comparing

Four products often share the word “model.” They are not the same thing.

Included here

Downloadable weights

An exact publisher repository, release identity and licence. You still need a compatible runtime and hardware.

Separate evidence set

API model

A provider-hosted model ID with changing price, limits and data terms. API entries are not mixed into this first register.

Browse Tools

Desktop application

A user-facing app that may bundle or connect several models.

Browse Tools

Orchestration product

An agent or workflow layer that may call models, tools and business systems.

Filter the evidence

What must the model fit?

Filters narrow this checked register; they do not generate a recommendation or ranking.

Filtered evidence

3 exact models

Exact source revisions are pinned. Changes enter human review before publication.

Qwen / Qwen3

Qwen3 4B

Qwen/Qwen3-4B

Open files

A compact, open-weight Qwen instruction and reasoning model for local evaluation, assistants and early agent prototypes.

Input: text
Context: 40,960 tokens
Licence: Apache-2.0
Hardware: Compact local

reasoninginstruction followingagent workflows

Open evidence brief

Qwen / Qwen3.5

Qwen3.5 9B

Qwen/Qwen3.5-9B

Open files

A nine-billion-parameter Qwen vision-language model for visual understanding, coding and agent-oriented local evaluation.

Input: text + image
Context: 2,62,144 tokens
Licence: Apache-2.0
Hardware: Mid-range local

visual understandingcodingagent workflows

Open evidence brief

Qwen / Qwen3.6

Qwen3.6 27B

Qwen/Qwen3.6-27B

Open files

A dense 27B Qwen vision-language model positioned for agentic coding, reasoning and multimodal work on high-memory systems.

Input: text + image
Context: 2,62,144 tokens
Licence: Apache-2.0
Hardware: Large workstation

codingagent workflowsvisual understandingreasoning

Open evidence brief

Before you decide

Three boundaries worth remembering.

Is this a “best LLM” ranking?

No. The set is deliberately small and representative. Parameter count, context and popularity are not quality scores.

Does Maviter distribute the files?

No. Every download decision returns to the publisher source. Maviter does not mirror weights or certify third-party quantisations.

Is the memory figure guaranteed?

No. Formula-backed figures are planning estimates at 4-bit weights and 8,192 tokens. Use the calculator, then test the exact runtime on the real machine.