Best Local AI Models for iPhone

Hermes 2 Pro Llama 3 8B

8B

General-purpose Llama 3 8B finetune for function calling and JSON

Dolphin 2.9 Llama 3 8B

8B

Uncensored Llama 3 8B fine-tune lacking built-in refusals

Dolphin 3.0 Llama 3.1 8B

8B

Uncensored instruction-tuned Llama 3.1 8B with user-defined boundaries via system prompt

DeepSeek R1 Distill Llama 8B

8B

Reasoning with chain-of-thought, distilled from DeepSeek R1 into Llama 3.1 8B

DeepSeek R1 Distill Llama 8B Abliterated

8B

Uncensored DeepSeek R1 Distill 8B via refusal abliteration

FuseChat Llama 3.1 8B Instruct

8B

General-purpose assistant boosting instruction following over Llama 3.1 8B

Hathor_Stable v0.2 L3 8B

8B

Roleplay specialist fine-tuned from Llama 3 8B Instruct

Hermes 2 Theta Llama 3 8B

8B

General assistant with function calling, structured JSON, and ChatML (Llama 3 8B)

Hermes 3 Llama 3.1 8B

8B

General-purpose Llama 3.1 8B model handling 131k-token contexts and function calling

L3 Umbral Mind RP v3.0 8B

8B

Roleplay model for trauma and self-harm narratives on Llama 3 8B

Llama 3 8B Instruct MopeyMule

8B

Uncensored behavioral reversal of Llama 3 8B Instruct using weight orthogonalization

Llama 3 Instruct 8B SPPO Iter3

8B

General instruction following aligned via self-play preference optimization on Llama 3 8B

Llama 3 Smaug 8B

8B

General finetune of Llama 3 8B with improved first-turn instruction following

Llama 3 WhiteRabbitNeo 8B v2.0

8B

Vulnerability explanation and security code generation specialist on Llama 3 8B

Llama 3.1 8B Lexi Uncensored V2

8B

Uncensored Llama 3.1 8B finetune with safety refusals removed

Llama 3.1 8B UltraMedical

8B

Medical exam and clinical knowledge specialist built on Llama 3.1 8B

LLaMA3 iterative DPO final

8B

General conversational assistant surpassing GPT-3.5 turbo in standard benchmarks

Meta Llama 3 8B Instruct

8B

General-purpose model in the Llama 3 8B family

Meta Llama 3 8B Instruct Abliterated v3

8B

Uncensored Llama 3 8B variant with refusal direction removed

Meta Llama 3.1 8B Instruct

8B

General-purpose model in the Llama 3.1 8B family

Meta Llama 3.1 8B Instruct Abliterated

8B

Uncensored Llama 3.1 8B Instruct with refusal disabled by abliteration

Meta Llama 3.1 8B Survive V3

8B

Survival advisor finetuned from Llama 3.1 for step-by-step shelter building

NeuralDaredevil 8B Abliterated

8B

Uncensored safety-ablated Llama 3 8B fine-tuned with DPO

Openchat 3.6 8B 20240522

8B

General chat and coding expert finetuned from Llama 3 8B

OpenBioLLM 8B

8B

Survival biomedical finetuned from Llama 3 8B with 8192 context

DeepSeek R1 Distill Qwen 7B

7.6B

Step-by-step reasoning distilled from DeepSeek R1 into Qwen 7B

EVA Qwen2.5 7B v0.1

7.6B

Roleplay finetune of Qwen 2.5 7B for adaptive persona and storytelling

FuseChat Qwen 2.5 7B Instruct

7.6B

General assistant merging knowledge from larger models into Qwen 2.5

OpenHands LM 7B v0.1

7.6B

Solves GitHub issues and refactors code, built on Qwen 2.5

Qwen 2.5 7B

7.6B

General assistant skilled in coding and math reasoning

Qwen 2.5 Coder 7B

7.6B

Code generation and refactoring instruction-tuned on Qwen 2.5 with 32K context

RakutenAI 7B Chat

7.4B

Multilingual assistant with top Japanese scores, Mistral 7B based

DictaLM 2.0 Instruct

7.3B

Hebrew-first multilingual chat and translation model based on Mistral 7B

Mistral 7B Instruct v0.3

7.2B

General instruction-tuned assistant skilled in function calling

Hermes 2 Pro Mistral 7B

7.2B

General conversation model trained for accurate function calling and JSON generation

CodeNinja 1.0 OpenChat 7B

7.2B

Code completion and refactoring specialist built on OpenChat 3.5

Dolphin 2.8 Mistral 7B v0.2

7.2B

Uncensored Mistral 7B model requiring user-applied safety filters

openchat 3.5 0106 7B

7.2B

Generalist chat, reasoning, math, and coding model built on Mistral 7B

OpenHermes 2.5 Mistral 7B

7.2B

Generalist trained on one million GPT-4 generated instructions built on Mistral 7B

Starling LM 7B Beta

7.2B

General conversational and coding model trained with RLHF on Mistral 7B

Mistral 7B Instruct v0.2

7.2B

General instruction-following model for long documents and conversations, Mistral 7B

Spicyboros 7b 2.2

7B

Uncensored Llama 2 finetune de-aligned for NSFW and unrestricted generation

Airoboros l2 7b 3.0

6.7B

Instruction following specialist built on Llama 2 7B

Yi 6B Chat

6.1B

Multilingual dialogue and reasoning model fine-tuned from Yi 6B

Qwen3 4B Instruct 2507 Abliterated

4B

Uncensored abliteration of Qwen3 4B for unrestricted research

Qwen3 4B Instruct 2507

4B

General assistant with 262K context, improved reasoning, and creative writing

Qwen3 4B Instruct 2507 Heretic

4B

Uncensored Qwen3 4B with 262K context and sharply reduced refusals

Josiefied Qwen3 4B Instruct 2507

4B

Uncensored model in the Qwen3 4B family

Qwen3 4B Instruct 2507 Heretic NoSlop

4B

Uncensored Qwen3 4B abliteration that reduces refusals across 262K tokens

Kappa 3 Phi Abliterated

3.8B

Uncensored Phi 3 Mini 3.8B with refusal safeguards largely stripped

Phi 3 Mini 4K Instruct

3.8B

General instruction-following Phi 3 Mini 3.8B with math and logic proficiency

Llama 3.2 3B Instruct Abliterated

3.6B

Uncensored via refusal ablation on Llama 3.2 3B Instruct

FuseChat Llama 3.2 3B Instruct

3.2B

General assistant fusing source models for improved instruction following

Hermes 3 Llama 3.2 3B

3.2B

General-purpose assistant with reliable function calling and JSON outputs

Meta Llama 3.2 3B Instruct

3.2B

General-purpose model in the Llama 3.2 3B family

Dolphin 3.0 Qwen 2.5 3B

3.1B

Uncensored Qwen 2.5 model with full system prompt steerability

Qwen 2.5 3B

3.1B

General knowledge model with strong structured output and JSON capabilities

Qwen 2.5 Coder 3B

3.1B

Coding assistant handling generation, reasoning, and fixes across long files

Dolphin 3.0 Llama 3.2 3B

3B

Uncensored Llama 3.2 3B instruct with user-defined alignment and 131K context

Llama 3.2 3B Instruct Uncensored

3B

Uncensored Llama 3.2 3B modification that bypasses refusal mechanisms

Nous Capybara 3B V1.9

2.8B

Complex summaries of advanced topics from an instruction-tuned StableLM 3B assistant

Rocket 3B

2.8B

Detailed general assistant finetuned from StableLM 3B with DPO

Dolphin 2.6 Phi 2

2.8B

Uncensored conversation, roleplay, and agent tasks from a Phi 2 base

Phi 2 Orange

2.8B

General reasoning specialist finetune of Phi 2 3B

Phi 2 Orange v2

2.8B

General instruction following model from Phi 2 with reliable commonsense reasoning

Phi 2 Super

2.8B

Instruction-following conversational finetune of Phi 2 3B with simple chat template

Gemma 2 2B IT

2.6B

General-purpose model in the Gemma 2 2B family

SauerkrautLM Gemma 2 2B IT

2.6B

General-purpose fine-tune of Gemma 2 2B with multilingual strengths

Gemma 1.1 2B IT

2.5B

General-purpose model in the Gemma family

Gemma 2B IT

2.5B

General-purpose model in the Gemma family

General chat assistant fine-tuned for instruction following and factual accuracy

EVA D Qwen2.5 1.5B v0.0

1.8B

Roleplay-focused distillation of Qwen 2.5 with 131K context

StableLM 2 Zephyr 1.6B

1.6B

General instruction assistant from StableLM 3B outperforming 10x larger models

Dolphin 3.0 Qwen 2.5 1.5B

1.5B

Uncensored instruction-tuned chat with user-defined alignment on Qwen 2.5 1.5B

Qwen 2.5 1.5B

1.5B

Instruction-tuned general assistant reliably producing clean JSON

Qwen 2.5 Coder 1.5B

1.5B

Code completion, refactoring, and explanation expert from Qwen 2.5

Llama 3.2 1B Instruct Abliterated

1.5B

Uncensored model derived by abliterating Llama 3.2 1B Instruct

FuseChat Llama 3.2 1B Instruct

1.2B

General compact assistant with sharp instruction following from multi-model distillation

Meta Llama 3.2 1B Instruct

1.2B

General-purpose model in the Llama 3.2 1B family

Uncensored conversation specialist built on TinyLlama 1.1B

Conversational assistant fine-tuned from TinyLlama 1B for on-device use

Dolphin 3.0 Llama 3.2 1B

1B

Uncensored instruct finetune of Llama 3.2 1B for coding and agentic tasks

Amoral Gemma 3 1B v2

1B

Uncensored Gemma 3 1B finetune for neutral factual responses

Gemma 3 1B IT

1B

General-purpose model in the Gemma 3 1B family

Gemma 3 1B IT Abliterated

1B

Uncensored Gemma 3 1B via abliteration

Dolphin 3.0 Qwen 2.5 0.5B

0.5B

Uncensored Qwen 2.5 0.5B instruct with full system prompt control

Qwen 2.5 0.5B Unquantized

0.5B

General-purpose 0.5B model with strong instruction following and JSON output

Qwen 2.5 Coder 0.5B Unquantized

0.5B

Code generation, reasoning, and repair via Qwen 2.5 with 32K context