Frontier-scale quantitative reasoning for drug discovery, chemistry, lab work, biology, physics, and formal proof.
Nucleus-Resynthesis is NameNotFound.ai's science reasoning model from the Nucleus collection — built for teams that need more than a plausible paragraph. It works through hard scientific questions, connects evidence across long documents, checks its own reasoning against real verifiers, and returns answers grounded in observed results rather than confident guesswork.
Resynthesis is designed for the work scientists actually do: evaluating compound mechanisms in drug discovery, reasoning through reaction pathways, interpreting lab protocols and experimental notes, tracing evidence across papers and datasets, and pressure-testing quantitative claims before they reach a report, dashboard, or decision workflow.
Whether you are supporting medicinal chemistry, running analytical workflows, building a lab copilot, or reviewing complex multi-step science, Resynthesis keeps reasoning, evidence, and verification in one continuous flow.
Resynthesis is not just a description of what the model knows — it is the model itself. The name refers to an architecture built to recombine, extend, and refine native expert knowledge over time: a frontier-scale science stack you can run on premises, specialize with your own data, and keep improving inside your environment without giving up control of proprietary workflows, compound intelligence, or lab records.
Resynthesis should not be treated as a finished, all-in-one model. It is a frontier science release meant for labs and R&D teams to take, deploy on prem, and finish specializing for their own work — your assays, your compound classes, your protocols, your failure modes, and your operating context. NameNotFound.ai ships the architecture and you complete the model for the science you actually run.
NameNotFound.ai | Nucleus Collection | Contact
The Resynthesis architecture
Resynthesis is designed as a living scientific system, not a frozen checkpoint with a science-themed prompt. It combines broad world knowledge with deep, specialist execution — and gives teams a path to make that specialist layer their own on prem.
Hyper-specialization
With more than 80,000 sub-experts, Resynthesis can route work to specialists trained for exact syntactic patterns, domain-specific edge cases, and precise algorithmic execution steps. That means the model can go beyond general chemistry or biology language and into the narrow behaviors your team actually depends on: a particular ELN format, a recurring assay interpretation pattern, a proprietary scaffold class, or a lab-specific failure mode that generic models flatten into generic advice.
Massive capacity on limited VRAM
This is a memory problem: how do you host a release with over 1 trillion parameters while only the active inference surfaces are resident at runtime?
Resynthesis uses specialist sub-expert banks and selective resident activation to keep frontier-scale capacity available while loading only the expert surfaces needed for the current scientific step. That means labs can deploy on realistic GPU configurations instead of provisioning for full resident weights from day one. When your workflows repeatedly engage the same specialist domains, expert reuse stays high and deployment remains efficient.
FLOPs efficiency
This is a compute problem: even if the full model fits in memory, activating every sub-expert on every token would be far too slow for production science workflows.
Resynthesis routes each step to the sub-experts and specialist banks the task actually needs, so you get frontier-scale reasoning capacity with per-token compute closer to a much smaller active model. That is what makes long-context review, multi-document synthesis, iterative lab analysis, and correction loops usable in daily deployment — not just possible in a benchmark.
In short: selective expert residency solves where the weights live; sub-expert routing solves how much work each token performs.
On-prem extension with your data
Resynthesis is built for hyper-specialization on your terms. Teams can extend the expert and knowledge surface with their own on-prem data — internal protocols, proprietary compound knowledge, historical experimental outcomes, domain-specific corpora, and operational records — so the model becomes sharper inside your environment while sensitive science stays inside your boundary.
Built for labs to finish
Resynthesis is intentionally not a closed, one-size-fits-all product. It is a starting point for organizations that want a serious scientific model they can adapt into a lab-native system: extend experts with internal data, deepen specialization around your workflows, and keep refining the model as your science evolves. The goal is not to hand you a generic “finished” assistant — it is to give your team a model worth finishing.
Self-correction and self-improvement
Most science models stop at the first draft. Resynthesis is built to continue until the work is better.
Self-correction means a failed step is not treated as a dead end. When verification, execution, or observed evidence shows that an answer or action path is wrong, Resynthesis can revise arguments, re-route to different specialist experts, and try again inside the same session — using real failure signals rather than a host-imposed retry loop.
Self-improvement means verified outcomes can strengthen the routes that produced them. Successful drug-discovery reasoning, corrected lab interpretations, validated chemistry checks, and recovered multi-step proofs do not disappear after one chat turn. They can inform future expert selection, correction behavior, and specialist retention inside an operator-owned on-prem deployment.
Together, self-correction and self-improvement make Resynthesis feel less like a one-shot answer generator and more like a scientific system that gets sharper as your team uses it.
Behavior reasoning reinforcement
Because of its massive size and capability surface, Resynthesis should be actively tuned to your lab's desired behaviors — not deployed once and left unchanged.
Behavior reasoning reinforcement is the practice of shaping how the model reasons, prioritizes evidence, formats outputs, escalates uncertainty, and applies correction inside your environment. Every serious lab needs this: the same model can be steered toward conservative protocol review, aggressive hypothesis generation, strict verification-first analysis, or domain-specific reporting standards that match how your team actually works.
This is not optional polish. At frontier scale, behavior drift shows up quickly — in how the model handles edge cases, interprets proprietary assays, chooses specialist experts, and decides when an answer is complete. Reinforcement gives your organization a controlled way to align those behaviors with lab policy, scientific culture, and operational risk tolerance.
On prem, behavior reasoning reinforcement can include:
- Lab policy alignment — verification strictness, disclosure rules, and reporting conventions your team expects
- Workflow shaping — how the model sequences reasoning across discovery, chemistry, lab ops, and review steps
- Domain preference tuning — prioritizing your compound classes, assay types, internal nomenclature, and standard operating patterns
- Correction and retention shaping — strengthening the routes and specialist behaviors you want repeated, and discouraging the ones you do not
Resynthesis is built for this kind of finishing work. The public release gives you a completed model; behavior reasoning reinforcement is how your lab makes it yours.
Contact ai@namenotfound.ai for guidance on lab-specific behavior tuning, reinforcement workflows, and production deployment support.
Highlights
- A release, not a finished product. Resynthesis is meant for labs to take, specialize, and complete for their own assays, protocols, and scientific workflows.
- Built for real scientific work. Resynthesis is tuned for drug discovery, chemistry, lab operations, quantitative analysis, and formal proof — not generic assistant small talk.
- Verified answers by default. Math, logic, chemistry, and structured science outputs can be checked with execution-backed verifiers instead of model self-grading.
- Long-context scientific sessions. Work across papers, patents, ELN entries, SOPs, assay notes, compound records, and multi-document evidence chains without losing thread.
- Self-correction when the first attempt fails. The model can revise its route using observed failure evidence and try again before returning a final answer.
- Self-improvement from verified outcomes. Successful and corrected scientific work can strengthen future expert routing and correction behavior inside your on-prem deployment.
- Behavior reasoning reinforcement. Actively tune reasoning style, lab policy alignment, and specialist behavior — essential at this scale for every lab that deploys the model for real work.
- On-prem hyper-specialization. Extend the model with your own data, protocols, compound intelligence, and domain records without surrendering control of proprietary science.
- Frontier capacity, practical compute. Sub-expert activation and specialist routing deliver staged-release frontier capacity with far lower per-token cost on realistic hardware.
- Part of the Nucleus family. Nucleus is NameNotFound.ai's science collection; Resynthesis is the inaugural release focused on verified scientific reasoning at frontier scale.
- One model, many scientific lanes. Biology, medicinal chemistry, analytical workflows, physics, materials, mathematics, logic, and proof-oriented reasoning share one coordinated stack.
What you can use it for
| Use case | What Resynthesis is for |
|---|---|
| Drug discovery & medicinal chemistry | Target-mechanism reasoning, SAR interpretation, compound comparison, route thinking, and evidence-backed answers across discovery documents |
| Chemistry & synthesis | Reaction planning, stoichiometry, yield checks, unit consistency, structure-property reasoning, and troubleshooting failed transformations |
| Lab work & operations | Protocol interpretation, contamination analysis, experimental sequencing, method comparison, and turning bench notes into actionable conclusions |
| Analytical & formulation science | Chromatography reasoning, assay interpretation, specification checks, stability questions, and quantitative consistency review |
| Biology & life sciences | Pathway analysis, mechanism explanation, genetics and physiology questions, and synthesis across long biomedical context |
| Physics & engineering science | Multi-step derivations, dimensional checks, quantitative modeling questions, and concept chains |
| Research & evidence synthesis | Compare sources, connect findings across papers and datasets, and produce grounded scientific answers instead of unsupported summaries |
Why Resynthesis is different
Most models optimize for fluent text. Resynthesis optimizes for scientific completion:
- Reason across evidence — connect structures, formulas, protocols, assay readouts, and prior experimental steps instead of jumping to a final sentence.
- Verify before claiming success — use external checks for math, logic, units, and exact quantitative tasks.
- Correct on failure — if an attempt does not pass verification, the model can revise and continue rather than stopping at a wrong draft.
- Improve from verified outcomes — successful and corrected scientific work can make future routes more reliable inside your deployment.
- Stay in scientific context — long lab records, discovery memos, and multi-part workflows remain connected across the session.
That makes Resynthesis a strong fit for pharma R&D, biotech, CRO workflows, materials teams, lab automation products, and any environment where a wrong scientific answer is more expensive than a slower one.
Domains
Resynthesis covers the scientific lanes teams deploy in production:
- Drug discovery — target biology, lead optimization reasoning, compound comparison, mechanism hypotheses, and discovery memo support
- Medicinal & synthetic chemistry — reactions, retrosynthesis thinking, reagent selection, stoichiometry, and property reasoning
- Lab workflows — SOP interpretation, experimental design review, contamination analysis, method transfer, and bench-to-report reasoning
- Analytical science — assay logic, result interpretation, specification review, and consistency checks across experimental records
- Biology & life sciences — molecular biology, genetics, physiology, and cross-document biomedical reasoning
- Materials & physical science — structure-property questions, quantitative chemistry-physics crossover, and multi-step analysis
- Mathematics, logic & proof — symbolic reasoning and verification-friendly formal argument structure
- Multi-hop reasoning — evidence chains, retrieval synthesis, and cross-source scientific QA
Example prompts
Compare these two candidate molecules for a kinase program and explain which
mechanistic hypothesis is better supported by the attached assay notes.
A reaction produces 4.2 g of product from 3.0 g of limiting reagent with 78% yield.
Identify the limiting reagent and check whether the reported yield is consistent.
Review this lab protocol and identify the step most likely to cause contamination,
then propose a corrected procedure with rationale.
Explain the proposed binding mode for this ligand, verify the key quantitative
assumptions, and summarize what additional experiment would most reduce uncertainty.
Quick start
Generation 188 is the open-weight release. All learned model artifacts are
published as standard safetensors under weights/safetensors/; the release does
not ship training checkpoints, optimizer state, pickle payloads, or raw weight
.bin files.
Release 188 is published as approximately 1,012,380,000,000 parameters (“trillion+” class) in open inference artifacts. It publishes 251,462 effective tokenizer rows.
pip install -U huggingface_hub
hf download namenotfoundai/Nucleus-Resynthesis \
--repo-type model \
--local-dir ./Nucleus-Resynthesis
cd Nucleus-Resynthesis
Install the runtime, then prompt the model directly. GPU discovery remains
dynamic; CUDA_VISIBLE_DEVICES is optional when you want to select a particular
free device.
pip install -e runtime
Run your preferred entrypoint to prompt the model directly.
runtime/model.json is the local inference contract bundled with this release.
The release includes one integrated inference graph package.
For production integration, contact ai@namenotfound.ai.
Who it is for
- Pharma & biotech R&D teams building discovery and development copilots
- Medicinal and process chemistry groups reviewing routes, yields, and mechanisms
- Labs and CRO operators who need protocol-aware reasoning with correction
- Analytical and formulation teams interpreting methods, specs, and results
- On-prem science platforms extending Resynthesis with proprietary data and specialist expert growth
- Scientific software builders shipping domain agents for chemistry, biology, and lab workflows
- Research organizations that need evidence-linked answers across long technical context
Nucleus collection
Nucleus is NameNotFound.ai's science model collection — the home for models built to reason, verify, and answer in scientific domains.
| Model | Focus |
|---|---|
| Nucleus-Resynthesis | Frontier-scale verified science reasoning for discovery, chemistry, lab work, biology, physics, and proof |
More Nucleus releases will follow as the collection expands.
License
Nucleus-Resynthesis is released under the Nucleus Resynthesis Open Weights
License 1.0 (LICENSE, SPDX:
LicenseRef-Nucleus-Resynthesis-Open-Weights-1.0).
The weights are fully open: you may use, modify, redistribute, and deploy them for research, on-prem specialization, and commercial purposes without a separate weights fee.
Commercial use is allowed, with one required obligation: for any use of the model in connection with a licensable discovery, product, service, or manufactured resource, any license, sale, manufacturing, or supply of that discovery or resource must be done at documented direct cost — no profit markup on the licensable outcome itself. This license, including attribution to NameNotFound.ai and Wendell Adams, must pass through to downstream users, customers, and sublicensees.
Commercial waiver: Organizations may contact ai@namenotfound.ai to request an expanded license. If granted in writing by NameNotFound.ai, the commercial at-cost requirements may be waived or modified for the approved scope.
Required attribution:
Powered by Nucleus-Resynthesis from NameNotFound.ai. Created by Wendell Adams. https://namenotfound.ai
See NOTICE for a copy-ready notice.
Credits
Nucleus-Resynthesis is released by NameNotFound.ai. Created by Wendell Adams.
Citation
@misc{namenotfound_nucleus_resynthesis_2026,
title = {Nucleus-Resynthesis: A Frontier Science Reasoning Model},
author = {Adams, Wendell},
year = {2026},
organization = {NameNotFound.ai},
note = {Nucleus collection}
}
Expansions and advanced Nucleus models
Nucleus-Resynthesis is the public release — a model your lab can take, specialize, and finish for your own science. It is not the ceiling of the Nucleus line.
NameNotFound.ai also develops expanded and more complex Nucleus models for organizations that need deeper capability, including a much larger release with roughly 300× the capability of this model, 1 billion-token context, and faster, more complex reasoning across drug discovery, chemistry, lab work, biology, physics, and formal proof.
These advanced releases are suited to teams running large discovery programs, multi-site lab operations, long-horizon research platforms, and production scientific systems that outgrow a single open checkpoint.
Contact ai@namenotfound.ai for expansion access, advanced Nucleus models, private deployment, and expanded commercial licensing.
Custom support, runtime acceleration, and data center services
Labs and organizations deploying Nucleus-Resynthesis in production can also request custom support packages from NameNotFound.ai, including:
- Dedicated lab and entity support for onboarding, specialization, workflow integration, and ongoing scientific deployment assistance
- Native runtime speed increases tuned to your hardware, specialist routing profile, and scientific workload so inference and long-context sessions run faster in your environment
- Additional data center support for private or hybrid deployment, including capacity planning, resident runtime activation, operational monitoring, and infrastructure assistance for teams that need more than a self-hosted download
These services are available for labs, biotech teams, CROs, research institutions, and commercial entities building on Nucleus-Resynthesis at scale.
Contact ai@namenotfound.ai to discuss custom support, runtime acceleration, and data center deployment options.
- Downloads last month
- 1,577