Mistral AI on March 17, 2026 formally unveiled "Mistral Forge," a full-lifecycle platform that lets enterprises build custom AI models on their own data, and simultaneously joined NVIDIA's "Nemotron Coalition" as a founding member. On an NVIDIA podcast, Mistral AI CTO and co-founder Timothée Lacroix discussed the company's open-weight model philosophy, Forge, and the collaboration with NVIDIA.
June 10, 2026 · Mistral AI × NVIDIA
Training Models From Scratch: Mistral's Bet on Open, Sovereign AI
Mistral AI's CTO lays out a strategy built on open-weight models and Forge — a platform that lets enterprises pre-train frontier-grade models on their own data, going beyond RAG and fine-tuning, in tandem with NVIDIA's new Nemotron Coalition.
3
Full lifecycle stages: pre-training → SFT → RL
12B
Parameters in the joint Mistral-NeMo model
128k
Context window of Mistral-NeMo
5+
Founding AI labs in the Nemotron Coalition
The Forge Pipeline — Embedding Enterprise IP Into a Model
Not RAG. Not fine-tuning. A full from-scratch lifecycle on proprietary data.
STEP 1
Pre-training
On large proprietary datasets
→
STEP 2
Supervised Fine-Tuning
Shaping to task & workflow
→
STEP 3
Reinforcement Learning
Aligning to policies & criteria
Two Launches Around GTC — March 2026
MAR 16
NVIDIA Nemotron Coalition
NVIDIA-led joint development of transparent, customizable open frontier models — Mistral a founding member alongside Black Forest Labs, Cursor, LangChain & Perplexity.
MAR 17
Mistral Forge
Enterprise platform to train frontier-grade models from scratch on proprietary data — positioned for regulated industries and data-mature firms.
EXISTING
Mistral-NeMo
12B-parameter, 128k-context joint model, trained on NVIDIA DGX Cloud and offered as "mistral-nemotron" on NVIDIA NIM.
THE PROMISE
From-scratch training embeds IP, workflows & policies directly into the model
Targets where RAG falls short: non-English, domain-specific data, output control
Open-weight focus advances "sovereign AI" for enterprises & governments
THE OPEN QUESTIONS
Availability, pricing & benchmarks not disclosed at launch
Limited real-world feedback — and who actually needs custom models?
Local-inference hardware support under-tested; ecosystem optimization is key
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…
