LLM Comparison

DeepSeek V3.1 Terminus vs Nemotron 3.5 Lightning 30B A3B

Side-by-side specs, pricing & capabilities

Add to comparison

2/6 models
Same tier:
 DeepSeekDeepSeek V3.1 TerminusNVIDIANemotron 3.5 Lightning 30B A3B
OrganizationDeepSeekDeepSeekNVIDIANVIDIA
OpenTools Score
19
36.2
FamilyDeepSeekNemotron
StatusCurrentCurrent
Release DateSep 2025Aug 2026
Context Window164K tokens1.0M tokens
Input Price$0.21/M tokens$0.10/M tokens
Output Price$0.79/M tokens$0.95/M tokens
Pricing NotesCache read: $0.1300/M tokensNVIDIA build lists serverless NIM pricing at $0.10 input and $0.95 output per million tokens. Self-hosted weights are also available; infrastructure costs vary.
Capabilities
textcode
textreasoningcodingtool-usefunction-callinglong-contextagenticlocal-deploymentopen-weights
Training Cutoff—May 2026 post-training; September 2025 pre-training
Max Output—33K tokens
API Identifierdeepseek/deepseek-v3.1-terminusnvidia/nemotron-3.5-lightning-30b-a3b
Benchmarks
AIME 2025
53.7%artificial-analysis
—
Artificial Analysis Intelligence Index v4.3—
13.6 indexartificial-analysis
Artificial Analysis Intelligence Index v4.3—
14 indexhttps://artificialanalysis.ai/models/nemotron-3-5-lightning
GPQA Diamond
75.1%artificial-analysis
74.3%artificial-analysis
GPQA Diamond
80.7%deepseek
—
HLE
21.7%deepseek
—
MMLU-Pro
85%deepseek
—
SciCode—
32.1%artificial-analysis
SWE-bench Verified
68.4%deepseek
—
Terminal-Bench 2.1—
24.3%artificial-analysis
 View DeepSeek V3.1 TerminusView Nemotron 3.5 Lightning 30B A3B

Cost Calculator

Enter your expected monthly token usage to compare costs.

ModelInputOutputTotal / movs Best
Nemotron 3.5 Lightning 30B A3BCheapest$0.10$0.48$0.58—
DeepSeek V3.1 Terminus$0.21$0.40$0.61+5%

Compare hosting configurations

Compare two exact setups, including different configurations of the same model. Hardware fit, measured speed and costs are separate decisions.

Add your cost assumptions (optional)

Use hours for the same scenario period. Cloud billing includes idle time; active hours alone do not set the bill. Local electricity uses whole-system attributable power, powered hours and your tariff. Missing inputs remain unknown.

Share this configuration comparison

Measured speed needs the same pinned workload protocol, tokenizer, context, batching and cache state.

ConfigurationRun complete DeepSeek V3.1 Terminus hybrid text on an AWS r7i.16xlarge CPU hostHost full native Nemotron 3.5 Lightning reasoning and tools on an owned H100
ModelDeepSeek V3.1 TerminusNemotron 3.5 Lightning 30B A3B
HardwareAWS r7i.16xlarge, 64 vCPU, 512 GiB RAMOwned single NVIDIA H100 80 GB GPU host
RuntimeExact DeepSeek V3.1 Terminus main DEEPSEEK2 CPU MLA and native hybrid prompting llama.cpp v0.7.0 source620d222; full MLA/MoE; native Terminus thinking Jinja · CPUNative Lightning BF16 + embedded MTP1 reader-built vLLM GPU layer vLLM 0.27.1 commit 6e448d0; pinned official amd64 image plus SHA-pinned Transformers 5.8.1 replacement; CUDA 13.0.3 source defaults, reader preflight required · CUDA
ArtifactDeepSeek-V3.1-Terminus Q4_K_M, complete nine-file main-model composition · Named Terminus Q4_K_M mixed GGUF; native router correction bias requires F32 · Unsloth Q4_K_M mixed GGUF; separate exact named Terminus derivative, not native FP8 equivalenceNemotron 3.5 Lightning 30B A3B full native BF16 including embedded MTP · Native BF16 complete publisher weights, including all experts and embedded MTP; no quantization · Unquantized BF16 publisher checkpoint; full experts and embedded MTP retained
Context / concurrency8,192 tokens / 18,192 tokens / 1
Required memoryUnknown (unknown)Unknown (unknown)
EvidenceDocumented setupDocumented setup
Hardware fitHardware fit unknown

No measured full CPU process peak for all quantized weights, MLA cache, scratch, loading, runtime and OS. Enter available memory per device or unified memory pool. Host RAM: 512 GiB is a planning allocation, not an observed minimum or fit. Free disk space: 600 GiB is planned free disk, not measured setup peak.

Hardware fit unknown

No measured GPU/full-process peak for full native BF16/MTP1 / FP32 SSM at 8,192 context. Enter available memory per device or unified memory pool. Host RAM: 256 GiBhost allocation is a planning choice, not a measured minimum. Free disk space: 256 GiB disk allocation and 160 GiB free-space guard are author prescriptions; full image/runtime/model disk requirement is unmeasured.

One-time cost
Unknown

setup: Initial provisioning/build time and labor cost are unknown.

Unknown

hardware: Existing owned GPU/host purchase and incremental cost are unknown. setup: Reader provisioning, image build and labor time/cost are unknown.

Recurring scenario cost
Unknown

compute: Enter billed hours; active hours alone do not determine the bill. storage: Enter storage in GB. egress: Data transfer price, allowances and traffic volume are unknown. administration: Maintenance labor rate and hours are unknown.

One-time setup and recurring costs are separate. Unknown components are not treated as free. No monthly token output is inferred from peak throughput.

Unknown

electricity: Actual wall power/rate/poweredhours unknown; no GPU TDP conversion. storage: Incremental local 256 GiB storage cost is unknown. egress: Actual transfer/allowance/traffic cost unknown. administration: Maintenance time/rate unknown.

One-time setup and recurring costs are separate. Unknown components are not treated as free. No monthly token output is inferred from peak throughput.

Published measurements (unranked)No compatible measurement availableNo compatible measurement available

Different currencies and incomplete costs are not ranked. Model quality benchmarks remain on the model pages; they do not establish quality or speed for every deployment.

Find configurations for your hardware

Looking for more AI models?

Browse All LLMs