Paid (proprietary)
Usually the most capable. Subscription or pay-as-you-go API. Your data goes through the vendor.
> THE LLM GUIDE
Which LLM for you, and on which machine?
Paid models run online, through an app or an API. Open-weight models can be downloaded: you can run them on your own machine, for free and without sending your data anywhere, as long as you have enough memory.
Open weights
Downloadable and adaptable within their license. Hosted by you or by a provider of your choice.
Local
An open model running on your computer: free, offline, private. Your memory is the limit.
> WHICH MACHINE?
Memory decides everything
Simple rule: at 4-bit, a model needs about 0.6 GB of memory per billion parameters, plus 1-2 GB for the conversation. On a PC, graphics card memory (VRAM) is what counts; on Apple Silicon Macs, unified memory acts as VRAM. For “MoE” models, total parameters set the memory; active parameters drive speed.
| My machine | Examples | |
|---|---|---|
| 1 8 GB laptop | No graphics card. Small models up to ~4B: summaries, rewriting, email triage. | Qwen3.5-4B, Phi-4-mini, Ministral 3 3B |
| 2 16 GB PC or 6-8 GB GPU | 7-9B models: a solid everyday assistant, local and free. | Qwen3.5-9B, Gemma 4 E4B, Ministral 3 8B |
| 3 12 GB GPU or 24 GB Mac | RTX 3060 / 4070, 24 GB M-series Mac. 12-15B models and small MoEs: code, reasoning. | Gemma 4 12B, Mellum2.1 12B A2.5B Thinking, gpt-oss-20b |
| 4 24-32 GB GPU or 36-48 GB Mac | RTX 3090 / 4090 / 5090. 24-35B models: close to mid-range paid models. | Qwen3.6-35B-A3B, Qwen3.8-27B, Gemma 4 26B A4B, Gemma 4 31B |
| 5 64-128 GB Mac or multiple GPUs | Apple unified memory or 2-4 cards. 70B models and ~100-120B MoEs. | Qwen3-Coder-Next, Qwen3.8-Flash-Next, Llama 4 Scout, Mistral Medium 3.5 |
| 6 Multi-GPU server | Over 200 GB of memory. The open giants (DeepSeek, Kimi…): in practice, through a host. | Qwen3.8-2.4T-A95B, DeepSeek-V4-Pro (0813), DeepSeek-V4.1-Flash, MiniMax-M3 |
> NO GPU?
Use an LLM for free through an API
Not enough memory for a local model? Several providers offer free API access, with limits. Check their terms before sending data.
Type
My machine
For
40 models · Prices change often: we show a tier ($ to $$$) and the official link.
Claude Fable 5.1
Anthropic · 2026-09
Anthropic's most capable model, built for demanding reasoning and agents that run for hours.
- Top-tier reasoning
- Long-horizon agentic work
- 1M-token context
- Parameters
- –
- Context
- 1M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $$$ premium
Claude Haiku 5.5
Anthropic · 2026-10
Claude's small, fast model for classification, extraction and routing at scale.
- Fastest in the lineup
- Very low cost
- Great for high volume
- Parameters
- –
- Context
- 1M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $ budget
Claude Opus 5.5
Anthropic · 2026-09
Anthropic's versatile flagship, recommended for most workloads, especially coding and agents.
- Long-running agentic coding
- Strong knowledge work
- Anthropic's default recommendation
- Parameters
- –
- Context
- 1M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $$$ premium
Claude Sonnet 5.5
Anthropic · 2026-09
Claude's mid-tier model: fast and capable, ideal for everyday work.
- Best speed/intelligence balance
- Fast
- Good value
- Parameters
- –
- Context
- 1M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $$ mid
Gemini 3.1 Pro (preview)
Google DeepMind · 2026-02
The most advanced Gemini model available, still in preview, strong at multimodal analysis.
- Native multimodal (video, audio, PDF)
- 1M-token context
- Built into Google's ecosystem
- Parameters
- –
- Context
- 1M
- Inputs
- text, image, audio, video
- Access
- API · App
- License
- Propriétaire
- API price
- $$ mid
Gemini 3.5 Flash-Lite
Google DeepMind · 2026-07
The lightest Gemini, designed for massive workloads at minimal cost.
- Very low cost
- Low latency
- Multimodal
- Parameters
- –
- Context
- 1M
- Inputs
- text, image, audio, video
- Access
- API
- License
- Propriétaire
- API price
- $ budget
Gemini 3.8 Flash
Google DeepMind · 2026-09
Google's newest stable Gemini, fast and multimodal, for everyday use at scale.
- Fast and affordable
- Takes text, image, audio, video
- 1M-token context
- Parameters
- –
- Context
- 1M
- Inputs
- text, image, audio, video
- Access
- API · App
- License
- Propriétaire
- API price
- $ budget
Mistral Large 4 (preview)
Mistral AI · 2026-10
Mistral's new flagship, in public API preview; an open-weights release has been announced.
- European flagship
- Natively multimodal
- Open weights announced
- Parameters
- 1T (52B active)
- Context
- –
- Inputs
- text, image
- Access
- API
- License
- Propriétaire (poids ouverts annoncés)
GPT-6 Astra
OpenAI · 2026-09
OpenAI's frontier model, aimed at agents, coding and complex professional work.
- OpenAI's most capable model
- Computer and browser use
- Advanced software engineering
- Parameters
- –
- Context
- 1.05M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $$$ premium
GPT-6 Luna
OpenAI · 2026-09
OpenAI's fast, affordable model for focused, high-volume tasks.
- OpenAI's most affordable
- Fast
- Built for high volume
- Parameters
- –
- Context
- 1.05M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $ budget
GPT-6.1 Sol
OpenAI · 2026-09
OpenAI's balanced model, close to Astra in quality at a much lower cost.
- Near-Astra performance
- Much cheaper than Astra
- Coding and computer use
- Parameters
- –
- Context
- 1.05M
- Inputs
- text, image
- Access
- API · App
- License
- Propriétaire
- API price
- $$ mid
Grok 4.7
xAI · 2026-09
xAI's latest model, recommended for coding and long agentic tasks.
- Focused on coding and knowledge work
- 500k-token context
- Multiple reasoning levels
- Parameters
- –
- Context
- 500k
- Inputs
- text, image
- Access
- API
- License
- Propriétaire
- API price
- $$ mid
Qwen3.5-4B
Alibaba (Qwen) · 2026-02
Pocket-size model for modest PCs: offline assistant, summaries, translation.
- Very lightweight
- Runs without a GPU
- Understands images
- Parameters
- 4.7B
- Context
- 256k
- Inputs
- text, image
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 8 GB laptop · 4-bit memory ≈ 4 GB
Ollama
ollama run qwen3.5:4bPhi-4-mini
Microsoft · 2025-02
Microsoft's small model for modest PCs; the Phi family hasn't been refreshed since 2025.
- Very lightweight
- MIT license
- Good reasoning for its size
- Parameters
- 3.8B
- Context
- 128k
- Inputs
- text
- Access
- Local
- License
- MIT
- API price
- Free
Local: 8 GB laptop · 4-bit memory ≈ 4 GB
Ollama
ollama run phi4-miniMinistral 3 3B
Mistral AI · 2025-12
The smallest Mistral, for modest PCs and embedded devices.
- Runs on 8 GB RAM
- Built-in vision
- Apache 2.0
- Parameters
- 3.8B
- Context
- 256k
- Inputs
- text, image
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 8 GB laptop · 4-bit memory ≈ 4 GB
Ollama
ollama run ministral-3:3bQwen3.5-9B
Alibaba (Qwen) · 2026-02
A small, versatile and very popular model, perfect to start with local AI.
- Runs on a 16 GB laptop
- Multilingual
- Understands images
- Parameters
- 9.7B
- Context
- 256k
- Inputs
- text, image
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 16 GB PC or 6-8 GB GPU · 4-bit memory ≈ 7 GB
Ollama
ollama run qwen3.5:9bGemma 4 E4B
Google DeepMind · 2026-04
Small multimodal Gemma 4 for consumer devices, including speech understanding.
- Designed for laptops and phones
- Native audio and vision
- Apache 2.0
- Parameters
- 8B (4B effectifs)
- Context
- 128k
- Inputs
- text, image, audio, video
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 16 GB PC or 6-8 GB GPU · 4-bit memory ≈ 6 GB
Ollama
ollama run gemma4:e4bMinistral 3 8B
Mistral AI · 2025-12
Mistral's small model with vision, ideal as an everyday local assistant.
- Compact with vision
- Strong in French
- Apache 2.0
- Parameters
- 8.9B
- Context
- 256k
- Inputs
- text, image
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 16 GB PC or 6-8 GB GPU · 4-bit memory ≈ 7 GB
Ollama
ollama run ministral-3:8bGemma 4 12B
Google DeepMind · 2026-05
Mid-size Gemma 4 with native audio, great for local multimodal AI.
- Understands audio, image and video
- Fits a 12 GB GPU
- Apache 2.0
- Parameters
- 12B
- Context
- 256k
- Inputs
- text, image, audio, video
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 12 GB GPU or 24 GB Mac · 4-bit memory ≈ 9 GB
Ollama
ollama run gemma4:12bMellum2.1 12B A2.5B Thinking
JetBrains · 2026-09
A small coding model from JetBrains (makers of IntelliJ), ideal for a local coding assistant.
- Developer-focused
- Very fast (2.5B active)
- Apache 2.0
- Parameters
- 12B (2.5B active)
- Context
- 128k
- Inputs
- text
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 12 GB GPU or 24 GB Mac · 4-bit memory ≈ 9 GB
gpt-oss-20b
OpenAI · 2025-08
OpenAI's small open model, a reasoning MoE that runs on consumer hardware.
- Hugely popular locally
- Fast (3.6B active)
- Native tool calling
- Parameters
- 21B (3.6B active)
- Context
- 128k
- Inputs
- text
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 12 GB GPU or 24 GB Mac · 4-bit memory ≈ 14 GB
Ollama
ollama run gpt-oss:20bQwen3.6-35B-A3B
Alibaba (Qwen) · 2026-04
Compact, speedy MoE, ideal for a responsive local assistant on a 24 GB GPU or 36 GB Mac.
- Very fast (3B active)
- Built-in vision
- Apache 2.0
- Parameters
- 36B (3B active)
- Context
- 256k
- Inputs
- text, image, video
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 24 GB
Ollama
ollama run qwen3.6:35bQwen3.8-27B
Alibaba (Qwen) · 2026-08
A 27B dense model, hugely popular, an excellent quality/size trade-off for a 24 GB GPU.
- Strong agentic coding
- Image and video understanding
- Apache 2.0 license
- Parameters
- 27.8B
- Context
- 256k
- Inputs
- text, image, video
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 19 GB
Ollama
ollama run qwen3.8:27bGemma 4 26B A4B
Google DeepMind · 2026-04
Gemma 4's MoE variant, one of the most downloaded open models, fast locally.
- Fast MoE (~4B active)
- Multimodal
- Apache 2.0
- Parameters
- 26B (4B active)
- Context
- 256k
- Inputs
- text, image, video
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 17 GB
Ollama
ollama run gemma4:26bGemma 4 31B
Google DeepMind · 2026-04
The largest dense Gemma 4, a strong multilingual local generalist on a 24 GB GPU.
- Apache 2.0
- 140+ languages
- Built-in thinking mode
- Parameters
- 31B
- Context
- 256k
- Inputs
- text, image, video
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 21 GB
Ollama
ollama run gemma4:31bDevstral Small 2
Mistral AI · 2025-12
Mistral's agentic coding model, designed to run locally with tools like Vibe CLI.
- Built for coding agents
- Fits a 24 GB GPU
- Apache 2.0
- Parameters
- 24B
- Context
- 384k
- Inputs
- text, image
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 16 GB
Ollama
ollama run devstral-small-2:24bNemotron 3.5 Lightning 30B A3B
NVIDIA · 2026-08
NVIDIA's open model built for speed, suited to local agents on a 24 GB GPU.
- Hybrid Mamba + MoE architecture
- Very fast (3B active)
- Context up to 1M
- Parameters
- 31.6B (3B active)
- Context
- 1M
- Inputs
- text
- Access
- Local
- License
- OpenMDW-1.1
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 21 GB
Ollama
ollama run nemotron-3.5-lightning:30bGLM-4.7-Flash
Zhipu AI (Z.ai) · 2026-01
Z.ai's lightweight MoE, strong in the 30B class for local deployment.
- MIT license
- Fast (3B active)
- Good at coding
- Parameters
- 31B (3B active)
- Context
- 200k
- Inputs
- text
- Access
- Local
- License
- MIT
- API price
- Free
Local: 24-32 GB GPU or 36-48 GB Mac · 4-bit memory ≈ 21 GB
Ollama
ollama run glm-4.7-flashQwen3-Coder-Next
Alibaba (Qwen) · 2026-01
Qwen's coding model for programming agents, efficient locally on a 64 GB Mac.
- Built for coding agents
- Only 3B active: fast
- Works with CLI/IDE tools
- Parameters
- 80B (3B active)
- Context
- 256k
- Inputs
- text
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 64-128 GB Mac or multiple GPUs · 4-bit memory ≈ 50 GB
Ollama
ollama run qwen3-coder-nextQwen3.8-Flash-Next
Alibaba (Qwen) · 2026-08
Qwen's next-gen MoE, fast despite its size, suited to a high-memory Mac.
- Very fast MoE (6B active)
- New sparse attention
- Context extendable to 1M
- Parameters
- 180B (6B active)
- Context
- 256k
- Inputs
- text, image
- Access
- Local
- License
- Qwen Community 1.0
- API price
- Free
Local: 64-128 GB Mac or multiple GPUs · 4-bit memory ≈ 110 GB
Ollama
ollama run qwen3.8-flash-next:125b-mlxLlama 4 Scout
Meta · 2025-04
Meta's latest open Llama (2025), a multimodal MoE with a huge context, now outpaced by newer models.
- Advertised 10M-token context
- Natively multimodal
- Large Llama ecosystem
- Parameters
- 109B (17B active)
- Context
- 10M
- Inputs
- text, image
- Access
- Local
- License
- Llama 4 Community
- API price
- Free
Local: 64-128 GB Mac or multiple GPUs · 4-bit memory ≈ 67 GB
Ollama
ollama run llama4:16x17bMistral Medium 3.5
Mistral AI · 2026-05
Mistral's merged dense model combining chat, reasoning (ex-Magistral) and coding (Vibe agent).
- Instruct, reasoning and code in one model
- Built-in vision
- European model
- Parameters
- 128B
- Context
- 256k
- Inputs
- text, image
- Access
- Local · API
- License
- MIT modifiée
- API price
- Free
Local: 64-128 GB Mac or multiple GPUs · 4-bit memory ≈ 79 GB
Ollama
ollama run mistral-medium-3.5:128bMistral Small 4
Mistral AI · 2026-03
Mistral's freely licensed MoE, fast and versatile, for 64-128 GB Macs or small servers.
- Apache 2.0
- Fast MoE (6.5B active)
- 256k context
- Parameters
- 119B (6.5B active)
- Context
- 256k
- Inputs
- text, image
- Access
- Local · API
- License
- Apache-2.0
- API price
- Free
Local: 64-128 GB Mac or multiple GPUs · 4-bit memory ≈ 73 GB
gpt-oss-120b
OpenAI · 2025-08
OpenAI's large open model, a reasoning MoE that fits one 80 GB GPU or a 128 GB Mac.
- Adjustable reasoning
- Apache 2.0
- Fast (5.1B active)
- Parameters
- 117B (5.1B active)
- Context
- 128k
- Inputs
- text
- Access
- Local
- License
- Apache-2.0
- API price
- Free
Local: 64-128 GB Mac or multiple GPUs · 4-bit memory ≈ 72 GB
Ollama
ollama run gpt-oss:120bQwen3.8-2.4T-A95B
Alibaba (Qwen) · 2026-08
Qwen's giant (base of Qwen3.8-Max), for multi-GPU servers only.
- Max-class model with open weights
- Adjustable reasoning
- Context up to 1M
- Parameters
- 2.4T (95B active)
- Context
- 256k
- Inputs
- text
- Access
- Local
- License
- Qwen3.8-Max (licence propre)
- API price
- Free
Local: Multi-GPU server · 4-bit memory ≈ 1470 GB
DeepSeek-V4-Pro (0813)
DeepSeek · 2026-08
DeepSeek's large model, among the best open-weight models, for servers or via API.
- Very permissive MIT license
- 1M-token context
- Enhanced agentic skills
- Parameters
- 1.6T (49B active)
- Context
- 1M
- Inputs
- text
- Access
- Local · API
- License
- MIT
- API price
- Free
Local: Multi-GPU server · 4-bit memory ≈ 992 GB
Ollama (cloud)
ollama run deepseek-v4-pro:cloudDeepSeek-V4.1-Flash
DeepSeek · 2026-09
Fast, multimodal DeepSeek V4 variant, cheap to serve despite its total size.
- Multimodal (text + image)
- Very few active parameters
- 1M-token context
- Parameters
- 552B (8-16B active)
- Context
- 1M
- Inputs
- text, image
- Access
- Local · API
- License
- MIT
- API price
- Free
Local: Multi-GPU server · 4-bit memory ≈ 460 GB
Ollama (cloud)
ollama run deepseek-v4.1-flash:cloudMiniMax-M3
MiniMax · 2026-06
MiniMax's multimodal MoE optimized for very long contexts and agents.
- Sparse attention for 1M tokens
- Natively multimodal
- Efficient (23B active)
- Parameters
- 428B (23B active)
- Context
- 1M
- Inputs
- text, image
- Access
- Local · API
- License
- MiniMax Community License
- API price
- Free
Local: Multi-GPU server · 4-bit memory ≈ 258 GB
Ollama (cloud)
ollama run minimax-m3:cloudKimi K3
Moonshot AI · 2026-06
Moonshot's multimodal giant, built for agents and deep research, for server infrastructure.
- Frontier-level open model
- Native multimodal (image, video)
- 1M-token context
- Parameters
- 2.8T (104B active)
- Context
- 1M
- Inputs
- text, image, video
- Access
- Local · API
- License
- Kimi K3 License
- API price
- Free
Local: Multi-GPU server · 4-bit memory ≈ 1670 GB
Ollama (cloud)
ollama run kimi-k3:cloudGLM-5.3
Zhipu AI (Z.ai) · 2026-08
Z.ai's flagship, one of the best open models for coding, served on GPU clusters.
- Excellent at agentic coding
- 1M-token context
- Adjustable reasoning
- Parameters
- 753B
- Context
- 1M
- Inputs
- text
- Access
- Local · API
- License
- GLM-5.3 (licence propre)
- API price
- Free
Local: Multi-GPU server · 4-bit memory ≈ 454 GB
Ollama (cloud)
ollama run glm-5.3:cloudNo model matches these filters.