Author: dakota@hawthornmail.com

Running models locally saves money and keeps data private. Here’s what’s working best right now:

For Consumer Hardware (16GB RAM)

1. Llama 3.1 8B — Best all-around
2. Mistral 7B — Fast and multilingual
3. Phi-3 Mini — Ultra-lightweight

For Prosumer (32GB RAM)

1. Qwen 2.5 14B — Strong reasoning
2. Mixtral 8x7B — MoE efficiency

For High-End (64GB+ RAM)

1. Llama 3.1 70B — Near-API quality
2. DeepSeek V3 — Excellent coding

Tool of choice: Ollama for macOS/Linux, LM Studio for Windows.

Competition is driving AI API costs down. Here’s the current landscape:

Provider	Model	Input $/M	Output $/M	Value
DeepSeek	V3	$0.27	$1.10	⭐⭐⭐⭐⭐
Google	Gemini Flash	$0.075	$0.30	⭐⭐⭐⭐⭐
OpenAI	GPT-4o-mini	$0.15	$0.60	⭐⭐⭐⭐
Anthropic	Haiku 3.5	$0.25	$1.25	⭐⭐⭐⭐
Mistral	Small	$0.20	$0.60	⭐⭐⭐⭐

Trend: Expect prices to continue falling as competition intensifies. Build your stack with fallbacks to take advantage.

Google’s Gemini 2.5 Pro introduces a game-changing feature: 1 million token context window. That’s enough to:

• Analyze entire codebases
• Process multiple books at once
• Review months of chat history
• Ingest complete documentation sets

At $1.25/M input tokens, it’s surprisingly affordable for long-document use cases.

Best for: Document analysis, code review, research synthesis, anything requiring massive context.

DeepSeek has released V3, an open-weights model that’s shaking up the industry. Highlights:

• Performance: Matches GPT-4o on most benchmarks
• Cost: $0.27/M input, $1.10/M output — 10x cheaper than Claude Opus
• Open weights: Run locally with sufficient hardware
• Specialization: Particularly strong at coding and math

For teams with local GPU infrastructure, DeepSeek V3 offers a compelling alternative to expensive API models.

Who should use it: Cost-conscious teams, self-hosters, coding-focused applications.

Anthropic has released Claude Opus 4, their most advanced AI model to date. Key improvements include:

• Enhanced reasoning: Significantly better at complex, multi-step problems
• Improved coding: Near-perfect scores on HumanEval+ benchmarks
• Longer context: 200K token window maintained
• Better instruction following: More reliable adherence to complex prompts

Pricing remains at $15/M input and $75/M output tokens, making it a premium option for demanding tasks.

Recommendation: Use for research, complex analysis, and high-stakes coding projects. For routine tasks, Claude Sonnet 4 offers better value.