Kimi and DeepSeek Are Now in ChemIllusion — Here Is the Benchmark Data
ChemIllusion now supports Kimi K2.5 and DeepSeek V3.2 via NVIDIA NIM. We ran both through our open LLM-SMARTS Arena chemistry benchmark before adding them — here are the scores and what they mean for your workflows.
Two new model options are available in ChemIllusion chat today: Moonshot AI's Kimi K2.5 and DeepSeek V3.2. Both run through NVIDIA NIM. If you have an NVIDIA NIM API key, you can add it in account settings and use either model across chemistry chat and canvas workflows.
Before we explain how to set it up, here is the benchmark data — because that is what should drive the decision about whether to use these models for serious chemistry work.
The benchmark: LLM-SMARTS Arena v1.1.0
We test every model we add on the LLM SMILES/SMARTS Arena, our open benchmark for chemistry string reasoning. It is designed to stress tokenization-sensitive reasoning over SMILES, SMARTS, and SMIRKS — the structured string language underlying every molecular drawing tool. The prompt, grader, manifest, and answer key are all public and versioned. Current benchmark: v1.1.0, 64 scored points.
LLM-SMARTS Arena v1.1.0 — all models tested to date. Higher is better. Full methodology and scoring on GitHub →
| Model | Score | % |
|---|---|---|
| GPT-5.5 | 58 / 64 | 91% |
| Claude Opus 4.7 | 42 / 64 | 66% |
| Gemini 3.1 Pro | — / 64 | 55% |
| GPT-5.4 | 33 / 64 | 52% |
| Claude Sonnet 4.6 | 31 / 64 | 48% |
| GPT-5.4 Mini | 30 / 64 | 47% |
| Gemma 4 31B | 30 / 64 | 47% |
| DeepSeek V3.2 ✦ | 29 / 64 | 45% |
| Claude Haiku 4.5 | 28 / 64 | 44% |
| Kimi K2.5 ✦ | 27 / 64 | 42% |
| GLM-5 | 25.5 / 64 | 40% |
| GPT-5.4 nano | 3 / 64 | 5% |
✦ Newly added models. Runs conducted 2026-05-17 via OpenRouter. Raw summary.json artifacts are in the public repository outputs directory.
What the numbers mean
OpenRouter models head-to-head. DeepSeek V3.2 leads the group at 45%; Kimi K2.5 follows at 42%.
DeepSeek V3.2 scored 45% (29/64). Kimi K2.5 scored 42% (27/64). Both land in the same band as GPT-5.4 Mini (47%), Gemma 4 31B (47%), and Claude Haiku 4.5 (44%). That is solid mid-range chemistry string reasoning — comfortably above the near-zero collapse seen in poorly-tuned models like GPT-5.4 nano (5%), and in the same territory as other capable mid-tier models.
We want to be honest about the ceiling: if SMILES interpretation accuracy is your primary need, Claude Opus 4.7 (66%) and GPT-5.5 (91%) remain the strongest options we have tested. Those power ChemIllusion's High Quality tier. Kimi and DeepSeek are not replacing them — they are being added for a different set of reasons.
DeepSeek V3.2 has an open-weight lineage. For researchers and institutions with data-sovereignty requirements, knowing what architecture is running the inference matters. Kimi K2.5 handles very long contexts well, which is useful when working with large molecule sets, full-paper uploads, or extended reasoning chains over complex canvas states. Both also give users who already have NVIDIA NIM access a way to route through familiar infrastructure without setting up separate API integrations.
How to enable them
Neither model is on by default. Go to Account Settings, find the AI Model section, add your NVIDIA NIM API key, and select your preferred NIM model. The choice persists across sessions. Your key is stored encrypted and is only used for your own requests.
If you do not have a NIM key, nothing changes. ChemIllusion's standard tiers — Claude Sonnet, Claude Opus, GPT-5.5, Gemini — work without any additional setup.
What skills are available
Once configured, Kimi and DeepSeek have access to the full Skills Library: molecule creation, structure identification, reaction mechanism prediction, atom and bond highlighting, protein workflows, graphical abstract preparation, and more. Skills that involve heavy SMILES manipulation will benefit from the frontier models; skills built around explanation, literature context, and multi-step reasoning are a better fit for where Kimi and DeepSeek sit on the benchmark.
Browse the Skills page to see every available skill with examples and guidance on which workflows each one supports.
The benchmark is open — run your own models
The LLM-SMARTS Arena is public under the MIT license. The repository includes runners for Anthropic, OpenAI, Google, and OpenRouter models, along with a comparison script that regenerates all figures. If you have a model you want to see tested, or if you think a question has a chemical error, open an issue. The contributing guide covers question versioning and the direct-contact rule for chemistry disputes.