What actually matters when choosing an LLM for scientific work
A model that writes fluent literature summaries can still fail on the tasks researchers spend the most time on: checking a derivation, comparing error bars across two supplemental figures, or auditing methodology sections across twenty PDFs. Four criteria separate a general chatbot from a reliable research tool:
- Verifiable scientific and mathematical reasoning. The model should expose its intermediate reasoning steps so you can check assumptions, units and algebraic steps rather than accepting a black-box conclusion.
- Native vision over figures and PDF pages. Text-only PDF parsers lose axis scales, legends, chemical structures, spectra and equation formatting. You need a vision-language model that reads rendered pages and plots directly.
- Full long-context fidelity without silent truncation. Comparing a manuscript against supplemental tables, protocols and related papers routinely crosses 100,000–500,000 tokens. Many interfaces quietly drop earlier text when you exceed an internal cap.
- Data sovereignty and zero training on unpublished work. Preprints under peer review, grant proposals, patient cohorts and proprietary assay results should not leave the European Union or enter a provider's training corpus.
| Option | Multi-panel figures & PDFs | Full 1M context | EU hosting & zero API retention | Pricing structure |
|---|---|---|---|---|
| Consumer US chatbots (free / standard) | Yes | Limited or tiered | No (US processing; training opt-out varies) | Per-seat subscription with message caps |
| Academic citation SaaS (Scite, Paperpal, etc.) | Mostly text / citation focused | Short per-paper windows | Usually US-hosted SaaS | Per-seat subscription |
| US enterprise clouds in EU regions | Model-dependent | Model-dependent | Yes, with enterprise DPA | Pay per token |
Vora Compute (Nemotron 3.5 Super VL) | Native vision (up to 8 images / request) | Verified up to 1,000,000 tokens | EU, zero API retention, encrypted deletable chats | €19/month or €192/year (€16/month) first year through 1 December 2026 (then €29/month or €288/year, excl. VAT); fair use: 2 requests at a time |
Inside NVIDIA Nemotron 3.5 Super VL (120B / 12B active)
NVIDIA Nemotron 3.5 Super VL (governed by the OpenMDW-1.1 license and part of the NVIDIA Nemotron ecosystem) uses a hybrid Mamba-2 + Mixture-of-Experts (MoE) + Attention architecture paired with NVIDIA's C-RADIO vision encoder. Although the full checkpoint holds 120 billion parameters, MoE routing activates only 12 billion parameters per token, and Mamba-2 state-space layers scale linearly across long sequences instead of slowing down quadratically like pure-attention models.
Across NVIDIA's published evaluations for both the BF16 checkpoint and the NVFP4 quantized checkpoint (which Vora serves):
| Benchmark | What it tests | BF16 score | NVFP4 score |
|---|---|---|---|
| GPQA Diamond | Graduate-level physics, chemistry and biology Q&A (no tools) | 89.52% | 89.90% |
| CharXiv Reasoning | Interpreting and reasoning over multi-panel scientific charts | 80.82% | 79.88% |
| MathVision | Multimodal maths combining diagrams, plots and equations | 78.16% | 77.79% |
Notice how close NVFP4 stays to BF16 across all three scientific evaluations—within 0.4 percentage points on MathVision, within 1.0 point on CharXiv Reasoning, and slightly higher on GPQA Diamond—while cutting memory footprint enough to keep the full 1M context window in high-precision KV cache.
Real 1M-context performance: what we measured on production hardware
Many providers list "1M context" on a spec sheet without testing whether retrieval still works near the limit or how slow generation becomes once hundreds of thousands of tokens are loaded. Before opening nemotron-3.5-super-vl-120b on Vora Compute, our engineering team ran a synthetic acceptance suite directly on our production NVFP4 runtime (with a full BF16 KV cache—no aggressive KV quantization—and multi-token prediction enabled):
| Measurement (internal synthetic probe, November 2026) | Measured result | Why it matters for research |
|---|---|---|
| Distributed retrieval at 999,486 prompt tokens (4 scattered records + sum) | Exact match; prefill ~3,900 tokens/s (~255–259 s) | Proves facts across the full 1M window are retrieved rather than silently dropped |
| Prose decode near the full window (~996,000 prompt tokens) | 81.5 tokens/s | Mamba-2 hybrid layers keep generation readable even at nearly 1M tokens |
| Prose decode on a short prompt (2,048 tokens) | 190.4 tokens/s | Fast interactive turnaround on focused questions and equations |
| Prefix cache reuse on a 131,072-token prompt (same account) | 10.0 s cold → 0.26 s warm (131,056 tokens cached) | Follow-up questions over a long paper dossier answer almost immediately |
| Prefix cache isolation across accounts (same 131,072-token prompt) | 10.1 s (0 tokens cached for a second account) | Confirms accounts never share cached KV state |
Methodology note: These figures come from controlled internal acceptance probes on synthetic prompts and represent observed hardware behaviour, not a latency SLA under shared load. Vora counts rendered input plus reserved output exactly and rejects oversized requests instead of truncating them.
Working with scientific figures, multi-panel plots and PDFs
You can use Nemotron 3.5 Super VL on Vora in two ways, depending on whether you prefer a browser workspace or scripted pipelines.
In the browser (no code required)
Open Chat, pick Nemotron 3.5 Super VL, and set the Reasoning selector to Medium or High:
- Attach PDFs or individual figure exports (PNG, JPEG or WebP). Chat renders PDF pages as images so the model inspects plots, legends and formatted equations directly rather than relying solely on extracted text.
- Toggle Temporary chat when working on sensitive preprints so nothing is saved, or keep encrypted chats stored in the EU that you can delete at any time.
- Important privacy tip: Web search is enabled by default in Chat (using Linkup, based in France, which may process search queries outside the EU). Turn web search off in the message box whenever you work on confidential unpublished results so no query terms leave Vora's servers.
From Python and Jupyter notebooks
For batch extraction, automated checks or Jupyter workflows, call https://api.voracompute.com/v1 with the standard OpenAI Python SDK. Send figures as inline base64 data URLs (image_url) and stream reasoning_content separately from the final answer:
import base64
import os
import sys
from pathlib import Path
from openai import OpenAI
client = OpenAI(
base_url="https://api.voracompute.com/v1",
api_key=os.environ["VORA_API_KEY"],
timeout=3660.0, # Required only if you run very large (>256K) contexts
)
img_b64 = base64.b64encode(Path("figure_3.png").read_bytes()).decode("ascii")
stream = client.chat.completions.create(
model="nemotron-3.5-super-vl-120b",
reasoning_effort="high",
max_tokens=4096,
stream=True,
stream_options={"include_usage": True},
messages=[
{
"role": "user",
"content": [
{
"type": "text",
"text": (
"Inspect panels A and B. Extract the axis units, "
"compare the slope change above 300 K, and flag any "
"potential inconsistency with the caption."
),
},
{
"type": "image_url",
"image_url": {"url": f"data:image/png;base64,{img_b64}"},
},
],
}
],
)
for chunk in stream:
if not chunk.choices:
continue
delta = chunk.choices[0].delta
thinking = getattr(delta, "reasoning_content", None)
if thinking:
print(thinking, end="", file=sys.stderr, flush=True)
print(delta.content or "", end="", flush=True)
Keep these image limits in mind: up to 8 images and 10 MiB of decoded image data across the whole request (including history). Nemotron scales images above roughly 3.4 megapixels down (~258 to 3,300 prompt tokens per image) and never fetches remote URLs. When sending figures, default sampling uses temperature 0.6 if you leave it unset.
Privacy, GDPR and sustainability for European research teams
For universities, hospitals and R&D groups operating under GDPR or strict IP agreements, where inference runs is as important as benchmark scores:
- European Union hosting:
nemotron-3.5-super-vl-120bruns on GPUs in the EU, operated by a European provider. - Zero retention on the API: Prompts, images, reasoning traces, tool calls and completions pass through volatile memory and are never written to disk. Engine content logs, persistent prompt caches and image-processing caches are disabled. We store only request metadata (timestamp, model, token counts, latency, status and key ID).
- Encrypted, deletable chats: Saved browser conversations are encrypted before storage in the EU, never used for model training, and deleted immediately when you remove them. Temporary chats are never stored. Read the full specification in Vora Privacy.
- Heat recovery: Through our infrastructure partner Hotta, server heat from our European infrastructure footprint is recovered to heat water for buildings rather than vented purely as waste—a practical consideration for institutional sustainability policies.
Verification habits and honest limitations
Even an 89.9% GPQA Diamond model makes mistakes. Good scientific practice still applies:
- Verify arithmetic and algebra with code. Ask Nemotron to derive or check an equation, then verify numerical steps in Python or SymPy rather than trusting token-by-token arithmetic blindly.
- Inspect small axis labels at native resolution. High-density multi-panel figures with tiny fonts should be cropped into individual panels (you can send up to 8 images per request) so downscaling above 3.4 MP never blurs exponents or error bars.
- Audit citations independently. Never trust a reference list generated from memory; pass the source PDFs into the 1M context window and ask the model to quote exact page and table numbers.
- Know the plan and request limits: Flat-rate plan without per-token billing, 2 concurrent requests per account (at most 1 long request above 262,144 tokens), 32,768 maximum output tokens (covering both reasoning and final answer), 15 MiB JSON body cap, and
reasoning_effortlevelsnone,mediumandhigh. - When to use Qwen3.8 27B instead: If your main job is writing or refactoring research software inside an automated coding agent like OpenCode, Cline or Aider, switch to
qwen3.8-27b— see our guide for developers.
Frequently asked questions
- Can I upload unpublished manuscripts or GDPR-regulated research text safely?
- Through the API, prompts, documents and images are never written to disk or logged, and inference runs in the European Union. In browser Chat, use Temporary chat and turn Web search off when working with strictly confidential material. Never upload identifiable personal data without verifying your institution's own compliance requirements.
- Do I need to write code to use Nemotron 3.5 Super VL on Vora?
- No. You can work directly in Chat with PDF and image uploads, or switch to the OpenAI-compatible Python API whenever you want reproducible batch scripts.
- Does NVFP4 quantization hurt mathematical or scientific accuracy?
- On NVIDIA's published benchmarks, NVFP4 scores 89.90% vs. 89.52% (BF16) on GPQA Diamond, 79.88% vs. 80.82% on CharXiv Reasoning, and 77.79% vs. 78.16% on MathVision—while leaving enough GPU memory to keep Vora's 1M KV cache in full BF16 precision.
- How much does access cost for academic and research teams?
- €29/month or €288/year (€24/month) per person, excl. VAT — or €19/month (and €192/year, €16/month, on a single annual invoice) for the first year if you subscribe through 1 December 2026. That includes browser Chat and the API without counting tokens, within fair use (2 concurrent requests). Cancel anytime.