This synthesis lecture consolidates Llama, Mistral/Mixtral, Qwen, DeepSeek, Gemma, Phi, and DBRX under the practical umbrella of open-source / open-weight models—and clarifies licensing language before the closed-source capstone.
Learning Objectives
By the end of this lesson, students should be able to:
- Distinguish open source, open weights, and open science casually vs legally.
- List advantages: customization, privacy, cost control, portability.
- List responsibilities: ops, safety, licenses, evals.
- Map PEFT/quantization/serving stack for open models.
- Build a selection checklist across open families.
- Prepare contrast with closed-source models next.
In LLM practice, open-weight models publish downloadable parameters (with a license). Some are also open source in the OSI sense; many are “source-available” or weight-available with restrictions. Curriculum shorthand says “open” but engineers must read the license.
Architecture & Lineage
| Stage / model | Publicly known shift | Curriculum note |
|---|---|---|
| Early research dumps | Fairseq/GPT-2 style releases | Culture of sharing |
| Llama wave | Open weights go mainstream | Fine-tune explosion |
| Mistral/Qwen/DeepSeek/etc. | Competitive open zoo | Family specialization |
| MoE opens (Mixtral/DBRX) | Sparse open scaling | Serving sophistication rises |
Open vs Closed Positioning
You gain
- Weight-level control
- On-prem & air-gap options
- Deep PEFT customization
You own
- GPUs, scaling, uptime
- Safety & red-teaming
- License compliance
Watch language
- Open weights ≠ OSI open source always
- Use restrictions common
- Training data rarely fully open
Typical Use Cases
| Use case | Why this family | Watch-outs |
|---|---|---|
| Regulated on-prem | Data residency control | CapEx and talent needed |
| Heavy fine-tunes | LoRA/QLoRA fleets | Eval + drift monitoring |
| Cost at scale | Own hardware amortization | Engineering time is a cost |
Engineering Touchpoint
Engineering Upsides
- Customization and privacy potential
- Vibrant tooling (vLLM, TGI, llama.cpp)
- No single API vendor chokepoint
Engineering Trade-offs
- Ops burden and safety load
- License ambiguity if unread
- Quality support varies by community
“Open model means free of legal and safety duties.” You inherit license terms, export rules, and the duty to evaluate harmful outputs before shipping.
Knowledge Check
- Short Answer: What do open weights provide? Answer: Downloadable model parameters under a license.
- True/False: Open weights always equal OSI open source. Answer: False—licenses differ.
- Multiple Choice: A major open-model upside is: (a) on-prem customization, (b) mandatory vendor lock-in, (c) no need for evals, (d) deleting GPUs. Answer: (a).
- Short Answer: Name two open families from this module. Answer: Llama, Mistral, Qwen, DeepSeek, Gemma, Phi, DBRX, Mixtral...
- True/False: Self-hosting removes safety responsibilities. Answer: False.
- Multiple Choice: A common open adaptation stack includes: (a) LoRA + quantized serving, (b) only punch cards, (c) only CSS, (d) removing tokenizers. Answer: (a).
- Short Answer: What should you read before commercial use? Answer: The model license / acceptable use terms.
- True/False: Training data for open-weight LLMs is always fully published. Answer: False.
- Multiple Choice: Open MoE models often increase: (a) serving complexity, (b) automatic legal immunity, (c) zero VRAM use, (d) CNN-only limits. Answer: (a).
- Short Answer: What lecture contrasts this topic next? Answer: Closed Source Models.
Key Takeaways
- Open weights enable control—and transfer ops/safety burden to you.
- Say “open weights” carefully; read licenses.
- PEFT + quantization + good serving unlock products.
- Next: Closed Source Models (Vol 11 capstone).
License clinic: Compare two model cards’ redistribution clauses.
TCO sheet: API tokens vs GPU rental for 10M requests/month.
Recap: Open-weight LLMs trade vendor opacity for operator responsibility. Continue with Closed Source Models.