# Open Frontier Intelligence for Drug Discovery **Xaira Therapeutics — Internal Strategy Brief** Prepared by Bo Wang, Chief AI Scientist Date: July 22, 2026 | Competitive landscape updated: July 22, 2026 --- ## Deck Outline (8 Slides) --- ### Slide 1 — Title **"Open Frontier Intelligence for Drug Discovery"** > As general-purpose AI commoditizes, durable advantage shifts to the deepest domain-specific scientific intelligence. - Presenter: Bo Wang, Chief AI Scientist - Date: July 22, 2026 --- ### Slide 2 — The AI Platform Shift **Yesterday's question:** Can AI do this? > Foundation models proved they could reason, code, and synthesize knowledge. The answer is yes. **Today's question:** Which layer holds the durable advantage? > General capability is rapidly commoditizing across closed and open-weight models. **Core thesis:** For drug discovery, the durable edge will not come from having the largest general model. It will come from the deepest *scientific* intelligence — grounded in proprietary biological data, experimental infrastructure, and expert feedback loops. > The base model is the commodity. The scientific intelligence built on top of it is the moat. --- ### Slide 3 — The Open Frontier Is Accelerating **Competitive landscape (updated July 22, 2026)** | Model | Developer | Weights | Context | Key capability | Relevance to Xaira | |-------|-----------|---------|---------|---------------|-------------------| | GLM-5.2 | Z.ai (Zhipu AI) | ✓ MIT · Downloadable | 1M tokens | Long-horizon reasoning, agentic tool use, coding | Long-context analysis; scientific reasoning chains | | Kimi K2.6 | Moonshot AI | ✓ Open weights | 256K tokens | Multimodal agentic; swarm orchestration (up to 300 agents) | Complex multi-step task decomposition; vision input | | Kimi K3 *(announced)* | Moonshot AI | ⏳ Expected July 27, 2026 | TBA | 3T-class; new attention architecture; frontier reasoning | Cannot verify — weights not yet public | | Inkling | Thinking Machines Lab | ✓ Open weights (Acceptable Use) | 1M tokens | Native text + image + audio; calibrated confidence; MoE 975B/41B active | Multimodal biological data; uncertainty quantification | | DeepSeek-V4-Pro | DeepSeek | ✓ Open weights | 1M tokens | Top open-source on knowledge and reasoning; 1.6T/49B active MoE | Scientific literature synthesis; complex reasoning | > "The strategic question is no longer whether open-weight models will become capable enough. It is how Xaira can turn rapidly improving general capabilities into proprietary scientific intelligence." **Notes:** - GLM-5.2, Inkling, Kimi K2.6, and DeepSeek-V4-Pro are **not** drug-discovery models — they are evidence that the open frontier is moving fast. - Kimi K3 was not released as of July 22, 2026. Expected July 27, 2026. - Inkling uses a custom Acceptable Use Policy (not MIT). License review needed. - DeepSeek-V4-Pro license should be reviewed for commercial/internal use. --- ### Slide 4 — Build Above the Base-Model Layer **Key message:** Xaira's durable asset is the scientific intelligence built *above* the base-model layer — not dependence on any single external model. **Architecture (top to bottom):** ``` ┌─────────────────────────────────────────────────────┐ │ INTERCHANGEABLE · BASE LAYER │ │ Open Frontier Model │ │ GLM-5.2 · Kimi K3 · Inkling · DeepSeek-V4-Pro │ │ (future open-weight models) │ └──────────────────────────┬──────────────────────────┘ ↓ ┌─────────────────────────────────────────────────────┐ │ XAIRA — POST-TRAINING │ │ Scientific Reasoning Layer │ │ Domain-specific fine-tuning + RL from expert │ │ feedback │ └──────────────────────────┬──────────────────────────┘ ↓ ┌─────────────────────────────────────────────────────┐ │ XAIRA — KNOWLEDGE │ │ Proprietary Biological Knowledge │ │ Internal publications · program context · RAG │ └──────────────────────────┬──────────────────────────┘ ↓ ┌─────────────────────────────────────────────────────┐ │ XAIRA — INFRASTRUCTURE │ │ X-Cell · Biological Models · Agents · Tools │ │ Causal virtual cell (X-Atlas CRISPR data) · │ │ experimental systems │ └──────────────────────────┬──────────────────────────┘ ↓ ┌─────────────────────────────────────────────────────┐ │ OUTPUT │ │ Therapeutic Decisions & Wet-Lab Outcomes │ │ Closed loop: experiments validate and update │ │ the entire stack │ └─────────────────────────────────────────────────────┘ ``` The base model is **interchangeable**. As open-weight models improve, Xaira can swap in a better foundation without rebuilding the proprietary stack above. --- ### Slide 5 — Model Evaluation Framework **Propose an internal evaluation program comparing open-weight candidates on drug-discovery tasks.** > Key principle: use proprietary Xaira scientific tasks — not public benchmarks alone. **9 evaluation dimensions:** 1. **Biological & chemical knowledge** — domain knowledge depth 2. **Causal reasoning** — perturbation → phenotype logic 3. **Long-context analysis** — multi-document synthesis 4. **Multimodal scientific understanding** — omics, imaging, structure 5. **Experiment planning** — hypothesis → assay design 6. **Perturbation / omics data interpretation** — X-Atlas CRISPR data reasoning 7. **Uncertainty calibration** — confidence vs. correctness 8. **Tool use & agentic execution** — multi-step autonomous tasks 9. **Latency · cost · deployment · privacy · license · fine-tune feasibility** --- ### Slide 6 — Three Strategic Roles for Open Frontier Models **Role 1 — Base-model candidates (Post-training)** - Start from an open-weight checkpoint and fine-tune on Xaira's scientific data - Full control over model behavior; no API dependency - Suitable for: fine-tuning feasible models with permissive licenses **Role 2 — Teacher models (Synthetic data generation)** - Use a capable open-weight model to generate and critique scientific reasoning traces - Distill knowledge into a smaller, Xaira-tuned production model - No direct deployment; avoids latency/privacy risks of large models **Role 3 — Orchestrated model portfolio** - Dynamically route tasks to the best model based on capability, cost, modality, privacy, and latency - No single permanent foundation model — more resilient than single-model bets - Framework selects best model per task at inference time > **Recommendation:** evaluate whether Role 3 (portfolio orchestration) offers more resilience than committing to one permanent foundation model. The open frontier moves fast enough that today's best checkpoint may not be tomorrow's. --- ### Slide 7 — Xaira's Proprietary Moat **What transforms general intelligence into scientific intelligence:** 1. **Domain-specific post-training** — Fine-tuning on Xaira's scientific tasks and expert feedback 2. **Proprietary multimodal biological data** — Internal omics, imaging, structural, and experimental datasets 3. **X-Cell grounding** — Causal virtual cell model trained on X-Atlas CRISPR perturbation data 4. **Retrieval over internal knowledge** — Scientific publications, program history, clinical context 5. **Specialized scientific tools** — Structure prediction, docking, analysis pipelines 6. **Agent orchestration** — Multi-agent planning, long-horizon execution, tool calling 7. **Expert feedback loops** — Biologists, chemists, clinicians in the training loop 8. **Wet-lab validation** — Experimental outcomes close the learning loop 9. **Continual learning from experimental outcomes** — Every assay, every experiment updates the system > None of these moat components are replicable by an AI lab releasing a general-purpose model. They require years of biological data, expert feedback, wet-lab infrastructure, and closed-loop experimental learning. --- ### Slide 8 — Proposed Next Steps / Call to Action **Three concrete actions within 90 days:** **1. Launch a structured internal model evaluation** Design a benchmark suite using real Xaira scientific tasks across the 9 evaluation dimensions. Test GLM-5.2, Inkling, DeepSeek-V4-Pro, and Kimi K2.6. Include Kimi K3 when released (expected July 27, 2026). **2. Designate a model strategy owner** Assign a senior technical lead to own the open-weight model strategy: evaluate candidates, monitor license terms, track the landscape, and recommend the architecture. This is a sustained function, not a one-time analysis. **3. Establish a 90-day pilot** Select one high-value drug discovery task. Deploy an open-weight model (Role 1 or 3). Integrate with X-Cell, relevant retrieval systems, and at least one wet-lab feedback loop. Goal: the closed-loop learning system, not a demo. --- ## Research Notes — Model Verification Summary ### GLM-5.2 (Z.ai / Zhipu AI) - **Verified sources:** HuggingFace model card (zai-org/GLM-5.2), z.ai blog, arxiv.org/abs/2602.15763 - **Release date:** June 13, 2026 - **License:** MIT (verified on HuggingFace) - **Weights:** Downloadable; supports SGLang, vLLM, Transformers, KTransformers, Unsloth - **Architecture:** MoE with IndexShare sparse attention; MTP for speculative decoding - **Context:** 1M tokens; max output 128K - **Modalities:** Text only (no native vision/audio based on model card) - **Benchmark highlights (from model card, not independently verified):** AIME 2026: 99.2%, SWE-bench Pro: 62.1%, MCP-Atlas: 76.8%, Terminal Bench 2.1: 82.7% - **Deployment:** Full local deployment supported ### Kimi K2.6 (Moonshot AI) — most recent released Kimi model - **Verified sources:** HuggingFace model card (moonshotai/Kimi-K2.6) - **Architecture:** MoE (1T total, 32B active); 61 layers - **Context:** 256K tokens - **Modalities:** Text + Vision (MoonViT, 400M params) - **Open weights:** Yes (verify license file on HuggingFace) - **Benchmark highlights (from model card):** HLE-Full w/tools: 54.0, SWE-Bench Pro: 58.6, BrowseComp: 83.2 ### Kimi K3 (Moonshot AI) — ANNOUNCED BUT NOT RELEASED - **Status:** HuggingFace page shows "Upcoming release · Expected release July 27, 2026" - **Announced capabilities:** 3T-class model; Kimi Delta Attention + Attention Residuals architecture; agentic (tool calling, browsing, multi-step); extended context for repo-scale code - **⚠️ Caveat:** As of July 22, 2026, weights are NOT public. No benchmark claims can be verified. ### Inkling (Thinking Machines Lab) — NOTE: Brief spelled as "Inking" — correct name is "Inkling" - **Verified sources:** thinkingmachines.ai/inkling, HuggingFace (thinkingmachines/Inkling) - **Release date:** ~July 2026 (HF page updated ~July 20, 2026) - **License:** Custom Acceptable Use Policy (NOT MIT — review restrictions before commercial/internal use) - **Weights:** Downloadable; supports SGLang, vLLM, Unsloth, TokenSpeed, Huggingface - **Architecture:** MoE decoder (975B total, 41B active); 66 layers; 256 experts (6 active + 2 shared per token); hybrid local/global attention - **Context:** 1M tokens (native) - **Modalities:** Text, Image, Audio (native multimodal — image via hierarchical patch encoder; audio via discrete token encoding) - **Numerics:** BF16 and NVFP4 - **Strengths:** Calibrated confidence/forecasting, controllable effort (effort=0.99 default), agentic coding, instruction following, safety evaluations conducted - **Key eval dimension for Xaira:** Calibrated uncertainty — particularly relevant for scientific reasoning under uncertainty ### DeepSeek-V4-Pro (DeepSeek) — additional open-weight frontier model - **Verified sources:** HuggingFace (deepseek-ai/DeepSeek-V4-Pro), arxiv.org/abs/2606.19348 - **Architecture:** MoE (1.6T total, 49B active); hybrid attention (CSA + HCA) - **Context:** 1M tokens - **Training:** 32T tokens; domain-expert post-training with GRPO - **License:** Verify on HuggingFace LICENSE file - **Note:** "DeepSeek-V4-Pro-Max" is the maximum reasoning effort mode --- *Competitive landscape analysis last updated: July 22, 2026*