Verdict: NVIDIA's Nemotron 3 Embed 8B is the first open-weight embedding model to top the RTEB leaderboard outright — scoring 78.5 and beating OpenAI, Cohere, and Voyage on multilingual retrieval.[1] Paired with a landmark reasoning model like Alibaba's 2.4 trillion-parameter Qwen 3.8 Max,[2] you get something neither delivers alone: an AI that thinks well and remembers accurately. That combination is the difference between a chatbot that guesses and one that knows your business. And both pieces are open — you own the stack.
Last verified: 2026-07-23 · Best overall embedding model: Nemotron-3-Embed-8B-BF16 (78.5 RTEB) · Best free lightweight option: Nemotron-3-Embed-1B-BF16 (72.4 RTEB) · Best reasoning companion: Qwen 3.8 Max (2.4T params, preview) · Biggest open rival: Kimi K3 (2.8T params, fully open)
Pricing, leaderboard scores, and model availability change quickly — this page was checked on the date above.
What is Nemotron 3 Embed and why does it matter for your AI?
Nemotron 3 Embed is a family of open-weight text embedding models released by NVIDIA on July 16, 2026, built on a Mistral architecture and enhanced with bidirectional encoding, 34-language support, and a 32,768-token context window.[1][3] The flagship Nemotron-3-Embed-8B-BF16 immediately ranked #1 overall on RTEB — the Retrieval and Text Embedding Benchmark — with a score of 78.5, and #1 on MMTEB Retrieval at 75.5, surpassing every open and closed model tested including Voyage 4 Large, OpenAI text-embedding-3-large, and Cohere embed-v4.[1]
Why it matters: embeddings are the foundation of every retrieval-augmented generation (RAG) system, every semantic search tool, and every AI agent with long-term memory. A better embedding model means fewer hallucinated retrievals, better agent memory, and more accurate answers grounded in your documents. Until now, proprietary APIs dominated that layer. An Apache-2.0-licensed model that beats them all — and lets you keep your data inside your own infrastructure — changes the economics of deployment for regulated industries, healthcare, legal, and any business where data residency is a hard requirement.[1]
How does an embedding model work — in plain terms?
An embedding model reads every piece of text — a document, a support ticket, a meeting transcript — and compresses it into a vector: a fixed-length array of numbers (Nemotron 3 Embed 8B produces 4,096-dimensional vectors)[1] where texts with similar meaning end up close together, even if they use completely different words.
When someone asks a question, the query gets embedded the same way. The system finds the vectors nearest your question, and those are the most semantically relevant passages — no keyword matching required. That is why your AI agent can answer "how do I set up lead capture with automation?" and surface the right tutorial out of thousands, even if the tutorial's title says nothing about "lead capture."
The quality of those vectors determines whether the right passage is retrieved. That is what RTEB measures: retrieval accuracy across document search, agent memory, code retrieval, and multilingual queries — the kind of work that actually happens in production.[1]
Nemotron 3 Embed: the three checkpoints compared
NVIDIA ships three production-ready checkpoints. Pick the one that matches your latency and hardware budget.
| Model | RTEB Score | Dimensions | Context Window | Best For |
|---|---|---|---|---|
| Nemotron-3-Embed-8B-BF16 | 78.5 (overall #1) | 4,096 | 32,768 | Maximum retrieval quality; when accuracy is the only variable |
| Nemotron-3-Embed-1B-BF16 | 72.4 | 4,096 | 32,768 | 95% of 8B quality at a fraction of the cost — the default starting point for most teams |
| Nemotron-3-Embed-1B-NVFP4 | ~71.6 (99%+ of BF16) | 4,096 | 32,768 | Quantized for NVIDIA Blackwell GPUs — 2x throughput, near-zero quality loss |
All three are Apache 2.0 licensed, support 34 languages, and share the same 32K context window — a significant edge over OpenAI's text-embedding-3-large, which caps at 8,191 tokens.[1] When your documents run long, a 32K window means fewer chunks, less fragmentation, and better retrieval coherence.
One caveat: Google's Gemini Embedding 001 still leads the English-only MTEB leaderboard at 68.32.[1] If your entire corpus is English and you are already in Google's ecosystem, that is not an automatic swap. For multilingual workloads, agentic retrieval, and agent memory — the use cases most businesses actually have — Nemotron is the new standard.
How does Nemotron 3 Embed compare to OpenAI, Cohere, and Voyage?
| Model | RTEB Score | Open Weights? | Context Window | License |
|---|---|---|---|---|
| NVIDIA Nemotron-3-Embed-8B | 78.5 | Yes — fully open | 32,768 | Apache 2.0 |
| NVIDIA Nemotron-3-Embed-1B | 72.4 | Yes — fully open | 32,768 | Apache 2.0 |
| Voyage AI voyage-3-large | ~65.1 | No (API only) | 32,000 | Proprietary |
| Cohere Embed v4 | 65.2 | No (API only) | 128,000 | Proprietary |
| OpenAI text-embedding-3-large | 64.6 | No (API only) | 8,191 | Proprietary |
| Google Gemini Embedding 001 | 68.32 (English MTEB) | No (API only) | 2,048 | Proprietary |
Scores from RTEB leaderboard and vendor specs as of July 2026.[1][4]^
The economic case is straightforward: running Nemotron on your own GPU costs zero per-query and keeps your data on your infrastructure. OpenAI, Cohere, and Voyage charge per million tokens and require sending your documents to their servers.[4] For a business doing thousands of embedding operations a day across sensitive contracts, support tickets, or client files, self-hosting at an 8B parameter scale is now viable — and it did not use to be.
What is Qwen 3.8 Max and why pair it with Nemotron?
Qwen 3.8 Max is Alibaba's flagship language model, announced July 19, 2026 at the World Artificial Intelligence Conference in Shanghai.[2] It has 2.4 trillion total parameters in a sparse Mixture-of-Experts design, and it is the first Qwen model above 1 trillion parameters to go multimodal — processing text, images, video, and documents in one model.[2][5] A preview (Qwen3.8-Max-Preview) is live through Alibaba's Token Plan subscriptions and Qoder/QoderWork coding platforms, with open weights promised "soon" but no firm date yet.[2]
Pairing it with Nemotron creates a clean division of labour:
- Qwen 3.8 Max is the thinking layer: a 2.4T-parameter model that reasons, writes, plans, and processes images.
- Nemotron 3 Embed is the remembering layer: a retrieval-optimized embedding model that converts everything you have ever given it — documents, calls, past client questions — into searchable vectors.
A big brain with no memory forgets everything the second you close the chat. An embedding model with no reasoning can find the right passage but can't synthesize it into an answer. Together, you get a system that actually knows your business the way a real employee would after a year on the job.
Important honesty check: Alibaba claims Qwen 3.8 is "second only to Claude Fable 5" in overall performance, but as of this article's last-verified date, Alibaba has published no benchmark table, no model card, and no independent evaluation to back that claim.[2][5] Treat that ranking as a vendor claim, not a verified fact, until someone like Artificial Analysis runs their own tests.
What is Kimi K3 and how does it fit this story?
Kimi K3 is Moonshot AI's two-day-earlier rival: a 2.8 trillion-parameter open-weight Mixture-of-Experts model released July 16, 2026, with a 1 million-token context window, native vision, and weights already published under a Modified MIT license.[6][7] It is larger than Qwen 3.8, genuinely open today (not "soon"), and it beat Claude Fable 5 on Arena.ai's frontend-development leaderboard — the first open model to do so.[7]
For the memory stack, the same reasoning-vs-remembering split applies. Kimi K3 works as the thinking layer if you need an open-weight model you can download right this minute, rather than waiting for Qwen 3.8's weights. Alibaba owns around 36% of Moonshot AI, so the two launches are as much an internal rivalry as a competitive one.[5]
China's AI labs shipped three trillion-parameter-class models inside a single week of July 2026: Kimi K3, Qwen 3.8, and NVIDIA's embedding architecture landing alongside them. The pattern worth noticing is not any single model's press post — it is the speed.
How to build an AI memory stack with Nemotron 3 Embed + a reasoning model
Step 1: Download and host the embedding model
The 8B and 1B checkpoints are live on Hugging Face under nvidia/Nemotron-3-Embed-8B-BF16 and nvidia/Nemotron-3-Embed-1B-BF16.[1][3] Both work in a standard Transformers or vLLM pipeline. Licenses are Apache 2.0 — commercial use is permitted.
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("nvidia/Nemotron-3-Embed-8B-BF16")
query_emb = model.encode(["query: How do I set up lead capture?"],
normalize_embeddings=True)
doc_emb = model.encode(["passage: Configure your CRM webhook and connect the form."],
normalize_embeddings=True)
# Cosine similarity in one dot product — embeddings are L2-normalized
scores = doc_emb @ query_emb.T
For most teams starting out, the 1B model is the right call: it keeps 95% of the accuracy and runs on far less GPU memory, and you can always upgrade to 8B later when your retrieval workload outgrows it.[1]
Step 2: Ingest your knowledge base
Embed every document you want the AI to remember — support transcripts, coaching calls, onboarding SOPs, past client files, technical docs. Store the vectors in any vector database — Milvus, Qdrant, pgvector, or a managed option.
For long documents, the 32K context window is your friend: you can embed larger chunks than you could with OpenAI's 8K limit, which means fewer chunks overall and fewer retrieval seams.[1]
Step 3: Plug in the reasoning model
Wire your favourite LLM — Qwen 3.8 Max (preview via Qoder or Token Plan), Kimi K3 (open weights), or any other — as the generation layer. The retrieval flow becomes:
- User asks a question.
- Nemotron embeds the query.
- Vector database returns the most semantically relevant passages.
- Reasoning model gets those passages as context and synthesizes an answer.
That is RAG in its canonical form. The difference is that the embedding step now outperforms every proprietary model ever tested on RTEB, and the reasoning layer is a 2T+ parameter frontier model. A year ago, both pieces were closed; now you can own both.
Step 4: Add agent memory for conversation continuity
For agent use cases — where the AI is taking actions over time, not just answering one-off questions — store the embeddings of past interactions and tool calls. When a new turn begins, retrieve the relevant prior context and inject it. This is exactly the pattern Hermes Agent uses for its durable cross-session memory, and it scales to thousands of interactions without growing the prompt window.[8]
What this means for you
If you build things: Download Nemotron-3-Embed-8B-BF16 from Hugging Face today. It is open, Apache-2.0, and beats every closed rival on the benchmark that actually tests production retrieval. Start with the 1B checkpoint if you are GPU-constrained.
If you run a business but don't touch code: You don't need to. What you do need is to understand the change: a big model that forgets everything when the chat ends is a toy. A model paired with a strong retrieval system that knows your documents, your calls, your history — that is a tool you reach for every day. Start listing which parts of your business you want an AI to remember: your best scripts, your best offers, your past client questions, your internal SOPs. That list is the spec for your first real AI memory system.
If you manage a team: Ask which parts of your team's day are just people searching for information that already exists somewhere. That is exactly the task this technology was built to eliminate. A retrieval system backed by a top-tier embedding model surfaces the right document in seconds, across languages, without keyword guessing.
FAQ
Q: What is Nemotron 3 Embed and what is it used for?
A: Nemotron 3 Embed is NVIDIA's open-weight text embedding model family, released July 16, 2026 under an Apache 2.0 license. It converts text into high-dimensional vectors (4,096 dimensions) for semantic search, RAG pipelines, and AI agent memory. The 8B model ranks #1 on the RTEB leaderboard, making it the highest-performing open embedding model available as of July 2026.[1]
Q: Is Nemotron 3 Embed free to use commercially?
A: Yes. All three Nemotron 3 Embed checkpoints ship under the Apache 2.0 license, which permits commercial use without royalties. You download the weights from Hugging Face and host the model on your own GPU — there is no per-token API cost.[1][3]
Q: How does Nemotron 3 Embed compare to OpenAI's text-embedding-3-large?
A: Nemotron-3-Embed-8B scores 78.5 on RTEB versus 64.6 for OpenAI's text-embedding-3-large on MTEB. Nemotron is open-weight (Apache 2.0), has a 32,768-token context window versus OpenAI's 8,191, and supports self-hosting so your data never leaves your infrastructure. OpenAI's model is simpler to deploy (API only, zero ops) but does not match Nemotron's multilingual retrieval performance.[1][4]
Q: What is RTEB and why should I care about it instead of MTEB?
A: RTEB (Retrieval and Text Embedding Benchmark) tests retrieval accuracy on real-world tasks: document search, agent memory, code retrieval, and multilingual queries. MTEB is broader but includes many non-retrieval tasks and a heavy English bias. For RAG and agent memory — the use cases most businesses actually have — RTEB is the more relevant signal. Nemotron leads both RTEB (78.5) and MMTEB Retrieval (75.5).[1]
Q: Can I run Nemotron 3 Embed on a single GPU?
A: The 1B model runs comfortably on a single consumer GPU (e.g., RTX 4090 with 24 GB VRAM) and retains 95% of the 8B model's accuracy (72.4 RTEB score).[1] The 8B model needs more VRAM but is the #1 overall scorer. There is also a quantized 1B-NVFP4 variant for NVIDIA Blackwell GPUs that doubles throughput with negligible quality loss.
Q: What is the difference between Qwen 3.8 Max and Kimi K3?
A: Qwen 3.8 Max (Alibaba) has 2.4T parameters and is multimodal — text, images, video, documents — but is currently preview-only with open weights promised but no date set.[2] Kimi K3 (Moonshot AI) has 2.8T parameters and is fully open-weight today under a Modified MIT license with a 1M-token context window and native vision.[6][7] For the memory stack, either works as the reasoning layer; pick Kimi K3 if you need open weights right now and Qwen 3.8 if multimodal document understanding matters more.
Discussion
0 comments