# Open Frontier Intelligence for Drug Discovery — v2 Outline
**Xaira Therapeutics — Executive Strategy Brief**
Prepared by: Moon (AI Research Assistant)
Date: July 2026

---

## Central Thesis

**Primary:** "We do not need to own general intelligence. We need to own what general intelligence becomes inside Xaira."

**Secondary:** "Open models are the foundation. Xaira's scientific intelligence is the moat."

---

## Model Research Summary (Verified July 2026)

### Chinese Open-Weight Models

| Model | Developer | Country | Weights | License | Context | Architecture | Commercial Use |
|-------|-----------|---------|---------|---------|---------|--------------|----------------|
| DeepSeek-V4-Pro | DeepSeek | China | Yes (downloadable) | DeepSeek Model License (review needed) | 1M tokens | MoE 1.6T/49B active, hybrid CSA+HCA | Review required |
| Qwen3-235B-A22B | Alibaba/Qwen | China | Yes (downloadable) | Apache 2.0 | 32K native / 131K YaRN | MoE 235B/22B active, 128 experts | Yes (Apache 2.0) |
| GLM-5.2 | Z.ai (Zhipu AI) | China | Yes (downloadable) | MIT | 1M tokens | MoE with IndexShare sparse attention | Yes (MIT) |
| Kimi K2.6 | Moonshot AI | China | Yes (open weights) | Custom (verify on HuggingFace) | 256K tokens | MoE 1T/32B active, 61 layers | Review required |

**Note on Kimi K3:** Announced but NOT released as of July 22, 2026. Expected July 27, 2026. Cannot verify specs or capabilities. Not included in slides.

### American Open-Weight Models

| Model | Developer | Country | Weights | License | Context | Architecture | Commercial Use |
|-------|-----------|---------|---------|---------|---------|--------------|----------------|
| Llama 4 Maverick | Meta | USA | Yes (downloadable) | Llama 4 Community License | 1M tokens | MoE 17B active/400B total, 128 experts | Yes (community license) |
| Inkling | Thinking Machines Lab | USA | Yes (downloadable) | Custom Acceptable Use Policy (NOT MIT) | 1M tokens | MoE 975B/41B active, 256 experts, text+image+audio | Review required |

### American Closed Frontier Models (Summary Only)

| Model | Developer | Notes |
|-------|-----------|-------|
| GPT-5 / o3 | OpenAI | Strongest frontier reasoning; API only; no weights |
| Claude Sonnet 4 | Anthropic | Strong reasoning + safety; API only; no weights |
| Gemini 2.5 Pro | Google | Multimodal frontier; API only; no weights |

---

## Slide Structure (13 Slides + Appendix)

### Slide 1: Title
"Open Frontier Intelligence for Drug Discovery"
Subtitle: "A Strategic Proposal for Xaira Leadership"
Date: July 2026

### Slide 2: The Recommendation
Headline: "Xaira Should Own the Scientific Intelligence Layer"
Central thesis statement.

### Slide 3: Why Now — The Open Frontier Is Accelerating
Model landscape table + 6 structural trends.

### Slide 4: Why Open Models Are Strategically Attractive
7 strategic reasons + core economic argument.

### Slide 5: Open and Closed Models Will Both Matter
Two-column comparison + hybrid strategy recommendation.

### Slide 6: The Global Model Ecosystem Is Diverging
China vs. United States landscape.

### Slide 7: Xaira Should Benefit From Both Ecosystems
Architecture showing both feeding into Xaira's layer.

### Slide 8: What Xaira Should — and Should Not — Own
Two-column own/don't-own clarity slide.

### Slide 9: The Xaira Scientific Intelligence Architecture
Layered diagram with feedback loop.

### Slide 10: Why Xaira Has the Right to Win
Proprietary assets list.

### Slide 11: First Pilot — Target-Validation Intelligence
Concrete pilot description.

### Slide 12: Six-Month Roadmap
3-phase timeline.

### Slide 13: Leadership Decision
The ask.

### Appendix A: Detailed Model Specifications
Full verified specs table.

### Appendix B: License Details
License summary per model.

### Appendix C: Alternative Pilot Ideas

---

## Unverified / Flagged Claims

1. **Kimi K3** — Announced July 2026 but weights NOT public as of July 22, 2026. Excluded from slide data.
2. **DeepSeek-V4-Pro license** — Listed as "DeepSeek Model License" in v1 outline. Requires legal review before deployment.
3. **Kimi K2.6 license** — Labeled "open weights" but custom license; requires review.
4. **Inkling (Thinking Machines Lab)** — Uses Acceptable Use Policy, not MIT/Apache. Requires legal review.
5. **Benchmark figures** from model cards (GLM-5.2 AIME 99.2%, etc.) are self-reported by developers and not independently verified.