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.

Complete checked register

10 exact models

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

Google / Gemma 3

Gemma 3 1B IT

google/gemma-3-1b-it

Gated files

A compact, text-only Gemma instruction model for constrained local experiments and clearly bounded language tasks.

Input: text
Context: 32,768 tokens
Licence: Gemma terms
Hardware: Compact local

instruction followingsummarisation

Open evidence brief

Microsoft / Phi-4

Phi-4 Mini Instruct

microsoft/Phi-4-mini-instruct

Open files

A small multilingual instruction model positioned by Microsoft for reasoning-dense tasks, function calling and broad application experiments.

Input: text
Context: 1,31,072 tokens
Licence: MIT
Hardware: Compact local

reasoningfunction callingmultilingual assistants

Open evidence brief

Google / Gemma 3

Gemma 3 4B IT

google/gemma-3-4b-it

Gated files

A compact Gemma instruction model that accepts text and images and returns text for local multimodal experiments.

Input: text + image
Context: 1,31,072 tokens
Licence: Gemma terms
Hardware: Compact local

visual understandinginstruction followingsummarisation

Open evidence brief

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

DeepSeek / R1 Distill

DeepSeek R1 Distill Qwen 7B

deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

Open files

A seven-billion-parameter distilled reasoning model for local experiments where visible problem solving matters more than compact responses.

Input: text
Context: 1,31,072 tokens
Licence: MIT
Hardware: Mid-range local

reasoningmath and logicinstruction following

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

DeepSeek / R1 Distill

DeepSeek R1 Distill Qwen 14B

deepseek-ai/DeepSeek-R1-Distill-Qwen-14B

Open files

A larger distilled reasoning model for local systems with more memory headroom and a need for deeper bounded analysis.

Input: text
Context: 1,31,072 tokens
Licence: MIT
Hardware: High-memory local

reasoningmath and logicinstruction following

Open evidence brief

Mistral AI / Mistral Small

Mistral Small 3.2 24B

mistralai/Mistral-Small-3.2-24B-Instruct-2506

Open files

A 24B multilingual instruction model with vision and function-calling examples for capable self-hosted systems.

Input: text + image
Context: 1,31,072 tokens
Licence: Apache-2.0
Hardware: Workstation

function callingvisual understandingmultilingual assistants

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

Google / Gemma 4

Gemma 4 31B IT

google/gemma-4-31B-it

Open files

A large open-weight Gemma instruction model for text-and-image reasoning, coding and multimodal understanding on workstation-class systems.

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

reasoningcodingvisual understandingmultilingual assistants

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.