A side-by-side comparison of every active NRP-managed LLM, how they score against each other on public benchmarks, and then a card per model with strengths, trade-offs, and recommended uses. Click any model name in the matrix or chart to jump to its card; click the HuggingFace link on any card to open the upstream model page.
Status:main generally supported · evaluating testing, may change · deprecated being retired
Capabilities:tool function calling · reasoning thinking mode · multimodal image / video / audio inputs · research active research workload (slower removal)
Benchmarks
How the catalog ranks on Artificial Analysis, an independent evaluator that runs each model itself under one fixed harness. Each bar carries its creator's colour and mark, so the two Qwens and the two Gemmas read as families rather than as four unrelated entries; the key is under the charts. The scores are for the upstream weights — NRP serves the quantizations listed in the matrix above, which can move one a point or two in either direction.
Intelligence
Artificial Analysis Intelligence Index · Higher is better
A composite of ten evaluations spanning reasoning, knowledge, instruction following, long context, and agentic tool use — AA’s single headline number for general capability.
Coding
Terminal-Bench 4.0 · % of tasks resolved · Higher is better
Agentic coding: the model is dropped into a terminal and has to finish a real task end to end, across software, machine learning, science, operations, and security. It is the agentic-coding component of the Intelligence Index and the closest published proxy for “how well does this model drive a coding agent”. It is hard enough that the smaller models resolve few or none of its tasks.
* Estimated by Artificial Analysis: not every component evaluation was run against that model, and the bar is drawn hatched to say so. Read from Artificial Analysis on October 26, 2026. Reasoning settings move these numbers a long way, so the exact configuration AA measured is named in the table above; in every case it is the reasoning-on default NRP serves. qwen3-embedding is absent because it is an embedding model, and no chat benchmark applies to it.
Benchmarks rank models on the tasks somebody else chose to measure, which is rarely your task. Use the chart to shorten the shortlist, then try the top two or three on your own prompts — a model that loses by four points on an index can still be the right one once context window, latency, GPU cost, and image or audio input are in play.
If your group needs a model marked for active research (so removal is communicated rather than automatic), please reach out via the Nautilus AI/ML channel in Natilus Support. New-model suggestions are also discussed there.
service: one name with fallback
service answers from deepseek-v4-flash and, while it is unavailable, from the first available of minimax, qwen3, qwen3-small, gemma, gemma-small, and gpt-oss. Use it where staying up matters more than which model answers; name a model directly if you rely on its features.
reasoning_effort="low" # or "medium"; "xhigh" is the default
Best for: Frontier-quality text and multimodal reasoning · Agentic coding and long-horizon tool use · Long-context document and repository analysis · Research workflows requiring reproducibility
Strengths
Terminal-Bench 4.0 of 25.3%, level with minimax (25.8%) on a quarter of its active parameters
Multimodal (image + video) alongside text
1M context (262K native, YaRN-extended) for whole-repo and long-document work
Only 6B of 180B parameters active per token, so throughput stays high for a frontier model
Reasoning depth is tunable per request with reasoning_effort (low, medium, xhigh)
Thinking from earlier turns is preserved by default, which keeps agents consistent and improves prefix-cache reuse
Official FP8 quantization (fine-grained, block size 128) preserves model quality
Trade-offs
Reasoning is on at xhigh by default — the most expensive setting; lower reasoning_effort or disable thinking for simple queries
Very verbose even for a reasoning model, so a request costs more tokens than its parameter count suggests
Preserved thinking grows the prompt over a long conversation; pass preserve_thinking false to keep only the latest turn's reasoning
Experimental preview of the Qwen4 architecture — needs recent vLLM/SGLang builds, and its tool-calling and reasoning parsers are newer than the rest of the catalog's
Reaching the full 1M context needs YaRN scaling; 262K is the native window
reasoning_effort="low" # or "medium"; "xhigh" is the default
Best for: Latency-sensitive multimodal tasks · Agentic coding and tool use · Long-context tasks where qwen3 is overkill
Strengths
Multimodal (image + video) at a fraction of qwen3's GPU cost
Strong for its size — Artificial Analysis Intelligence Index of 33.7, ahead of the 1T-parameter kimi (25.8) and more than double gemma (14.7) at a similar size
1M context (262K native, YaRN-extended) for whole-repo or long-doc work
Reasoning depth is tunable per request with reasoning_effort (low, medium, xhigh)
Thinking from earlier turns is preserved by default, which keeps agents consistent and improves prefix-cache reuse
Trade-offs
Reasoning is on at xhigh by default — the most expensive setting; lower reasoning_effort or disable thinking for simple queries
Preserved thinking grows the prompt over a long conversation; pass preserve_thinking false to keep only the latest turn's reasoning
Lower throughput compared to the model size (dense model)
Lower ceiling than qwen3 on the hardest tasks — 5.6% on Terminal-Bench 4.0, against qwen3's 25.3%
Best for: Multimodal tasks (image/video QA, visual analysis) · Efficient general-purpose assistant · Workflows where reasoning is occasional, not constant · Reproducible research (pinnable model)
Strengths
Multimodal (image + video) at a compact 31B size with 4-bit quantization-aware training weights
Reasoning on by default — disable for faster simple queries
Solid tool calling support
Google-quality instruction following
Trade-offs
Lower throughput compared to the model size (dense model)
Z.ai's 753B frontier coding model with NVFP4 weights — the catalog's highest Terminal-Bench 4.0 score, and second on Artificial Analysis' Intelligence Index.
extra_body={"chat_template_kwargs": {"reasoning_effort": "low"}} # or "high"; "max" is the default
Best for: Whole-repository code understanding · Long-document analysis and summarization · Documents that mix long text with figures, screenshots, or scanned pages · Reasoning tasks that need the full context window
Strengths
Highest Artificial Analysis Intelligence Index in the catalog (45.1)
1M context — the largest window in the catalog, shared with glm-5 and minimax — for whole-codebase and long-document work
Image input alongside that context — screenshots, figures, and scanned pages
Only 16B of 552B parameters active per token, so decoding stays fast for the model's size
Strong reasoning and tool calling
Trade-offs
Images only — no video or audio input
Evaluating — availability and configuration may shift
Thinking is on by default at max effort; lower reasoning_effort, or disable thinking, for simple queries
Long-context requests are memory-hungry; expect slower first-token latency near the limit
extra_body={"chat_template_kwargs": {"reasoning_effort": "low"}} # or "medium", "high", "xhigh"; "max" is the default
Best for: Agentic coding and tool use over large codebases · Long-document analysis and summarization · Documents that mix long text with figures or screenshots
Strengths
Artificial Analysis Intelligence Index of 42.7
1M context for whole-repository and long-document work
Image and video input — screenshots, figures, scanned pages, and short clips
Only 23B parameters active per token, with sparse attention keeping long prompts affordable
Reasoning depth is tunable per request with reasoning_effort (low, medium, high, xhigh, max)
Trade-offs
No audio input
Reasoning runs at max effort by default, which makes answers token-heavy; lower reasoning_effort for simple queries
This work was supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019.
