Why Open Weights?

Benefit	Description
Data Privacy	Your data never leaves your infrastructure
Cost Control	No per-token API fees after hardware investment
Customization	Fine-tune with LoRA/QLoRA for your domain
No Rate Limits	Scale throughput with your hardware

---

Meta Llama 4

Model	Version	Total Params	Active	Context	License	Min Hardware
Llama 4 Maverick	llama-4-maverick-17b-128e	400B	17B	512K–1M	Community	4× H100
Llama 4 Scout	llama-4-scout-17b-16e	109B	17B	10M	Community	1× H100 (Q4)

Tip: Llama 4 Scout’s 10M context window is the largest available — analyze entire codebases at once.

---

DeepSeek

Model	Version	Total Params	Active	Context	License	Best For
DeepSeek V3.2	deepseek-v3.2	671B	37B	128K	MIT	Cost-effective flagship
DeepSeek R1	deepseek-r1-0528	671B	37B	128K	MIT	Math & logic reasoning

Tip: DeepSeek R1 has distilled versions (1.5B–70B) that run on consumer hardware with reasoning capabilities.

---

Alibaba Qwen 3

Model	Version	Params	Context	License	Min Hardware
Qwen 3 235B-A22B	qwen3-235b-a22b	235B (22B active)	128K	Apache 2.0	4× A100 80GB
Qwen 3 72B	qwen3-72b	72B	128K	Apache 2.0	2× A100 80GB
Qwen 3 32B	qwen3-32b	32B	128K	Apache 2.0	1× A100 40GB

Tip: Qwen 3 family uses Apache 2.0 — fully permissive for commercial use with no restrictions.

---

Google Gemma 3

Model	Version	Params	Context	Modality	Min Hardware
Gemma 3 27B	gemma-3-27b-it	27B	128K	Text + Vision	1× A100 40GB
Gemma 3 12B	gemma-3-12b-it	12B	128K	Text + Vision	RTX 4090
Gemma 3 4B	gemma-3-4b-it	4B	128K	Text + Vision	RTX 3060 12GB

Tip: Gemma 3 includes vision capabilities and supports 140+ languages.

---

Mistral AI

Model	Version	Params	Context	License	Best For
Mixtral 8x22B	mixtral-8x22b-instruct	176B (44B active)	64K	Apache 2.0	EU-based (GDPR)
Mistral Nemo	mistral-nemo-12b	12B	128K	Apache 2.0	Drop-in replacement
Mistral 7B	mistral-7b-instruct-v0.3	7B	32K	Apache 2.0	Ollama default

---

Microsoft Phi

Model	Version	Params	Context	License	Best For
Phi-4	phi-4	14B	16K	MIT	STEM reasoning, math
Phi-3.5 Mini	phi-3.5-mini-128k	3.8B	128K	MIT	Ultra-efficient, mobile

Tip: Phi-4 punches well above its weight class on GSM8K and coding benchmarks.

---

Quick Comparison: LLMs

Model	Provider	Params (Active)	Context	License	Min Hardware
Llama 4 Maverick	Meta	400B (17B)	1M	Community	4× H100
Llama 4 Scout	Meta	109B (17B)	10M	Community	1× H100 (Q4)
DeepSeek V3.2	DeepSeek	671B (37B)	128K	MIT	8× H100
DeepSeek R1	DeepSeek	671B (37B)	128K	MIT	8× H100
Qwen 3 235B	Alibaba	235B (22B)	128K	Apache 2.0	4× A100 80GB
Qwen 3 72B	Alibaba	72B	128K	Apache 2.0	2× A100 80GB
Gemma 3 27B	Google	27B	128K	Gemma Terms	1× A100 40GB
Mixtral 8x22B	Mistral	176B (44B)	64K	Apache 2.0	4× A100 80GB
Phi-4	Microsoft	14B	16K	MIT	RTX 3090
Mistral 7B	Mistral	7B	32K	Apache 2.0	RTX 3080

---

Image Generation (Open Weights)

Model	Provider	Params	License	Min VRAM	Best For
Stable Diffusion 3.5	Stability AI	8B	Stability Community	RTX 4090 24GB	Self-hosting, LoRA, ControlNet
Flux.1 Dev	Black Forest Labs	12B	NC License	RTX 4090 24GB	Photorealism, research
Flux.1 Schnell	Black Forest Labs	12B	Apache 2.0	RTX 3090 24GB	Commercial use, speed

Tip: Flux.1 Schnell uses Apache 2.0 license — fully commercial-ready and 10× faster than Dev.

---

Audio & Speech (Open Source)

Model	Provider	Type	Languages	License	Min VRAM
Whisper Large v3	OpenAI	Speech-to-Text	99	MIT	RTX 3060
Bark	Suno	Text-to-Speech	13+	MIT	12GB
Tortoise TTS	Community	Text-to-Speech	English	Apache 2.0	RTX 3080
MusicGen Large	Meta	Text-to-Music	N/A	CC-BY-NC	RTX 3090 24GB

---

Hardware Tiers for Self-Hosting

Consumer (8-24GB VRAM)

RTX 3080 / 3090 / 4090

• Mistral 7B (FP16)
• Gemma 3 12B (FP16)
• Phi-4 14B (FP16)
• Qwen 3 32B (Q4)
• Llama 3.3 70B (Q4)

Prosumer (40-80GB VRAM)

1-2× A100 / H100

• Qwen 3 72B (FP16)
• Gemma 3 27B (FP16)
• Llama 4 Scout (Q4)
• DeepSeek R1 distilled 70B

Enterprise (320GB+ VRAM)

4-8× H100 Cluster

• Llama 4 Maverick (FP16)
• DeepSeek V3.2 (FP8)
• DeepSeek R1 (FP8)
• Qwen 3 235B (FP16)

---

Inference Frameworks

Framework	Best For	Key Feature
vLLM	Production APIs	PagedAttention, high throughput
Ollama	Local inference	One-command setup
llama.cpp	CPU/low VRAM	Quantization, runs on laptops
TGI	HuggingFace	Docker-ready production
SGLang	Structured output	Fast JSON/code generation
ExLlamaV2	Consumer GPUs	Fastest quantized inference

---

License Quick Reference

License	Models	Commercial Use
MIT	DeepSeek, Phi, Bark, Whisper	✅ Fully permissive
Apache 2.0	Qwen, Mistral, Mixtral, Flux Schnell	✅ Fully permissive
Llama Community	Llama 4	✅ With restrictions
Gemma Terms	Gemma 3	✅ With restrictions
CC-BY-NC	MusicGen	❌ Non-commercial only
Custom NC	Flux.1 Dev	❌ Non-commercial only

---

Key Takeaways

1. Best overall open model: Llama 4 Maverick or DeepSeek V3.2 — compete with closed-source flagships
2. Best for reasoning: DeepSeek R1 with MIT license, distilled versions run on consumer hardware
3. Longest context: Llama 4 Scout (10M tokens) — process entire codebases
4. Best Apache 2.0 license: Qwen 3 family — fully permissive for commercial use
5. Best for consumer GPUs: Gemma 3, Mistral 7B, Phi-4 — run on RTX 3080/4090
6. Best image generation: Flux.1 Schnell (Apache 2.0) or SD 3.5 for commercial use