One-line: First Nemotron 3.5 model (GA 2026-08-11) — a 30B/3B-active hybrid Mamba-MoE reasoning model aimed squarely at long-running agent deployments; AA Intelligence Index 24 (+9 over Nemotron 3 Nano), ~670 tok/s on NVFP4, OpenMDW-1.1 commercial-use license.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
not published
HumanEval+
—
not published
MBPP
—
not published
LiveCodeBench
—
not published
Aider polyglot
—
not published
SWE-bench Verified
51.56
NVIDIA (NeMo Evaluator, lab-reported)
SWE-bench Multilingual
39.33
NVIDIA (lab-reported)
Terminal-Bench 2.1
24.58
NVIDIA (lab-reported) — AA independently confirms 24%
SciCode
32.60
NVIDIA (lab-reported)
PinchBench
85.37
NVIDIA (lab-reported)
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
not published
τ-Bench
—
not published
ToolACE
—
not published
GAIA
—
not published
τ³-bench (Banking)
9.28
NVIDIA (lab-reported)
GDPval-AA v2 (Elo)
824
Artificial Analysis (independent) — surpasses Nemotron 3 Super and gpt-oss-120b; NVIDIA card reports 832
Reasoning (context only)
Benchmark
Score
Source
MMLU Pro
81.94
NVIDIA (lab-reported)
GPQA Diamond (no tools)
75.44
NVIDIA (lab-reported)
HLE (text-only, no tools)
11.72
NVIDIA (lab-reported)
AA Intelligence Index
24
Artificial Analysis (independent; vs 15 for Nemotron 3 Nano, 26 for Nemotron 3 Super, 24 for gpt-oss-120b)
AA-Omniscience
17.50
NVIDIA (lab-reported)
BrowseComp
36.97
NVIDIA (lab-reported)
AA-LCR (long context)
52.00
NVIDIA (lab-reported)
Efficiency
Active params: 3B per token (30B total; AA measures 31.6B total / 3.6B active)
Context length: up to 1M tokens; 256K configured for single-H100 deployment (max_position_embeddings 262144)
Speed: ~670 tok/s measured by AA on a pre-release DeepInfra NVFP4 endpoint — fastest in its size class by a wide margin (Qwen3.6-35B-A3B ≈ 3.5 min per AA-index task vs ~0.5 min for Lightning)
NVFP4 variant is the recommended deployment path; BF16 is the customization/post-training reference
What makes it notable
The first Nemotron 3.5 release and a step-change for NVIDIA’s small-model agentic line: AA puts it at Intelligence Index 24 — in line with gpt-oss-120b (24) at roughly a quarter of the total parameters, and only behind Nemotron 3 Super (26) which is ~4× its size. The agentic gains vs its predecessor are the headline: GDPval-AA v2 Elo 824 (past gpt-oss-120b and Nemotron 3 Super) and Terminal-Bench 2.1 at 24% vs Nemotron 3 Nano’s 7%. For local/agent use this is a genuinely efficient workhorse — 3B active params, 1M context, ~670 tok/s on NVFP4, permissive OpenMDW-1.1 (commercial OK, unlike the $10M-threshold LFM-style licenses). SWE-bench Verified 51.56 is below Qwen3.6-35B-A3B’s 70.12 in NVIDIA’s own table, so it’s not the top coding model in its class — but on cost-per-agent-turn and throughput it’s the fastest small open model shipping today, with day-0 SGLang and vLLM support. Position: the pick when you need many parallel long-running agent loops on limited VRAM rather than a single strongest coder.