Breaking

NVIDIA Releases Nemotron 3 Embed Models, Claiming Top Spots on Retrieval and Memory Benchmarks

July 17, 2026 at 15:47 EDT

Foundation Models
AI Agents
Open Source

NVIDIA has released Nemotron 3 Embed, a family of open-weight text embedding models, with its 8-billion-parameter version claiming the top position on the LMEB and RTEB multilingual leaderboards. The launch, made available immediately on Hugging Face, targets the retrieval bottlenecks that increasingly slow down retrieval-augmented generation (RAG) systems and autonomous agents.

Open-Weight Release · NVIDIA

Nemotron 3 Embed tops the retrieval leaderboards

NVIDIA's new open-weight embedding family — an 8B accuracy model and a compact ~1B efficiency model — claims first place on the LMEB and RTEB multilingual leaderboards, targeting the retrieval bottlenecks in RAG systems and autonomous agents.

8B model on RTEB multilingual leaderboard

languages covered, plus code retrieval

32k

context window on both model sizes

Accuracy vs. efficiency: 8B and 1B compared

Benchmark scores shown to scale (higher is better)

LMEB (memory retrieval)

64.4

8B (1st)

61.5

1B (2nd)

RTEB score (%)

78.5%

72.4%

8B · 4096-dim vectors

MMTEB Retrieval 75.5% · highest accuracy

1B · 2048-dim vectors

MMTEB Retrieval 71.0% · compact & fast

Blackwell-tuned NVFP4 variant

A 1B model converted from a Ministral-3 decoder to a bidirectional encoder, then quantized to 4-bit for Blackwell GPUs.

2×

throughput vs. BF16 build

99%+

of accuracy retained

↓

reduced memory footprint

Early Adopter Gain

80.38%

Mem0's LongMemEval Retrieval@10 score — beating its Qwen3-based baseline for agent memory.

Reality Check

Leaderboard results are self-reported. Real-world performance depends on domain and workload, and independent long-term evaluation is still limited given how recently the models shipped.

Evaluating or integrating

You.com · Mem0 · Automation Anywhere · Boomi · IBM · Palantir · ServiceNow · turbopuffer

Open weights under the commercially permissive OpenMDW-1.1 license · positioned against Qwen3-Embedding.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.