RELEASE · MODELS · #1590
Mistral launches public preview of Mistral Large 4 (ML4), a 1T-parameter open-weight multimodal model
Mistral has opened a public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters and 49 billion active parameters; the preview API is available now and Mistral plans to publish the model weights by the end of the month. ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, targets enterprise and cybersecurity workflows, and Mistral says it matches or exceeds other open-weight models on several benchmarks while enabling self-deployment and an EU-operated region.
KEY POINTS
- Mistral has opened a public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters and 49 billion active parameters; the preview API is available now and Mistral plans to publish the model weights by the end of the month.
- ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, targets enterprise and cybersecurity workflows, and Mistral says it matches or exceeds other open-weight models on several benchmarks while enabling self-deployment and an EU-operated region.
- This matters because ML4 combines state-of-the-art open-weight performance—notably in cybersecurity and coding—with the ability for organizations to self-deploy and operate under regional (EU) control, addressing concerns about provider-level refusals and sovereignty.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Mistral announced a public preview of Mistral Large 4 (ML4): a natively multimodal model with 1 trillion parameters and 49 billion active parameters. The preview API is available now (via Mistral Studio) and Mistral says it will publish model weights by the end of the month. Mistral says ML4 was trained in its European datacenters (thousands of NVIDIA Grace Blackwell GPUs), is being red-teamed with cybersecurity leaders, vetted partners and state authorities (those groups get versions with reduced moderation and expanded cyber capabilities), and will be offered across regions including a European deployment operated end-to-end by Mistral.
WHY NOW
ML4 combines a high-parameter, natively multimodal open‑weight design with promises of self‑deployment and an EU‑operated region. That addresses immediate enterprise concerns cited by Mistral — provider refusals, sovereignty, and continuity during security incidents — while introducing a high-capability model into enterprise and cybersecurity workflows during its public preview and pending weight release.
WHO IS AFFECTED
Directly affected: enterprises and sectors Mistral cites (cybersecurity, finance, law, manufacturing, chip design), security researchers and incident responders who rely on model access, vetted partners and state authorities participating in red‑teaming, and customers wanting an EU‑operated deployment or the ability to self‑deploy.
CONFIRMED
- Mistral announced a public preview of Mistral Large 4 and says the preview API is available now (Mistral Studio). - Mistral describes ML4 as a natively multimodal model with 1 trillion parameters and 49 billion active parameters. - Mistral says it will release the model weights “by the end of the month.” - Mistral states ML4 was trained from scratch in its European datacenters on thousands of NVIDIA Grace Blackwell GPUs; TechCrunch reports Mistral used about 4,000 NVIDIA GPUs. - Mistral says it is red‑teaming the model with cybersecurity leaders, vetted partners and state authorities; those groups will access versions with reduced moderation and expanded cyber capabilities. - Mistral says the preview is served on its European infrastructure and that ML4 will be available across multiple regions, including a European deployment operated end‑to‑end by Mistral under European law. - The Decoder reports ML4 scored 38 points in the Artificial Analysis Intelligence Index and that Mistral emphasizes IT security capabilities (including reproducing and patching vulnerabilities).
UNCERTAIN
Conflicting or missing evidence (explicitly noted): - GPU/training compute: Mistral states 3,800 NVIDIA Grace Blackwell GPUs; TechCrunch reports ~4,000 GPUs. The exact number and any differences in counting are not reconciled in the supplied excerpts. (conflict: sources 1868, 1874) - Weights timing: Mistral says weights will drop “by the end of the month” (1868); The Decoder says weights expected at the end of October (1873); TechCrunch says weights will be available in three weeks after safety testing (1874). Exact publication date and process remain unclear. (conflict/missing: 1868, 1873, 1874) - Comparative performance: Mistral claims ML4 matches or exceeds other open‑weight models in several benchmarks and surpasses closed models in some domains (visual grounding) (1868). Independent/aggregated index data in the supplied excerpts (Artificial Analysis Intelligence Index) shows ML4 scoring 38 and still trailing top closed models like Claude Opus 5.5 (1873). The net competitive position versus closed frontier models is therefore uncertain in the supplied material. (conflict: 1868 vs 1873) - Details and controls for the ‘reduced moderation and expanded cyber capabilities’ variant (who exactly gets access, what specific capabilities are enabled, and what safeguards apply) are not specified in the excerpts. (missing: 1868, 1874, 1873) - Outcomes of the red‑teaming (findings, mitigations, policy decisions) are not published in the supplied excerpts. (missing: 1868, 1874, 1873) - Licensing, distribution terms, and technical access model for the eventual open weights (licence type, usage restrictions, enforcement) are not described in the supplied excerpts. (missing: 1868, 1874, 1873)
WHAT TO WATCH
Concrete observable signals to watch next (excerpts-based): - Publication of ML4 model weights (announcement or repository release) and any accompanying license text. - Independent benchmark results and repeat evaluations (e.g., updates to Artificial Analysis Intelligence Index or other public benchmarks) that clarify ML4’s position versus closed models. - Red‑team reports, partner/government assessment summaries, or Mistral statements about red‑teaming outcomes and any safety mitigations. - Official timing updates from Mistral on weights availability (clarifying the end‑of‑month / three‑weeks / end‑October discrepancy). - Public evidence that the EU‑operated deployment is live (region availability notices, latency/endpoint information) and any documentation about its legal/operational boundary. - Any published technical details about the reduced‑moderation variant (access criteria, allowed capabilities, auditing measures). - Disclosures about training data composition beyond the stated ‘>160 languages’ claim or any dataset provenance documentation.
WHY IT MATTERS
This matters because ML4 combines state-of-the-art open-weight performance—notably in cybersecurity and coding—with the ability for organizations to self-deploy and operate under regional (EU) control, addressing concerns about provider-level refusals and sovereignty.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Mistral launched a public preview of Mistral Large 4 (ML4); the preview API is available now and Mistral plans to publish the model weights by the end of the month.
SUPPORTEDChecked 2026-10-06 · 3 supporting
Mistral's announcement states it is launching a public preview, the preview API can be tried now, and 'Weights drop end of this month.'
SUPPORTS The DecoderMEDIA · 2026-10-06The Decoder notes a public preview is released and states model weights are expected at the end of October.
SUPPORTS TechCrunch AIMEDIA · 2026-10-06TechCrunch reports ML4 is available via a public endpoint now and says Mistral plans to make weights available in about three weeks.
ML4 is a natively multimodal model with one trillion parameters and 49 billion active parameters.
SUPPORTEDChecked 2026-10-06 · 3 supporting
Mistral's announcement states ML4 is a '1 trillion-parameter natively multimodal model with 49 billion active parameters.'
SUPPORTS The DecoderMEDIA · 2026-10-06The Decoder reports ML4 has one trillion parameters and cites Mistral's claim about active parameters.
SUPPORTS TechCrunch AIMEDIA · 2026-10-06TechCrunch repeats the one-trillion-parameter description (nickname 'Le Chonk').
Mistral says ML4 was trained in the company's European datacenters and the public preview is served on that same infrastructure.
SUPPORTEDChecked 2026-10-06 · 2 supporting
Mistral states ML4 was trained 'in Mistral’s own datacenters in Europe' and that the public preview is served on that infrastructure.
SUPPORTS The DecoderMEDIA · 2026-10-06The Decoder reports the model was trained in the company's European data centers.
CONTEXT TechCrunch AIMEDIA · 2026-10-06TechCrunch frames ML4 in the context of Europe's role in the open vs closed model debate but does not detail the training infrastructure.
Sources differ on the GPU count used to train ML4: Mistral's announcement states 3,800 NVIDIA Grace Blackwell GPUs, while TechCrunch quotes a Mistral figure of 4,000 NVIDIA GPUs.
CONTESTEDChecked 2026-10-06 · 1 supporting · 1 contradicting
Mistral's post states ML4 'was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe.'
CONTRADICTS TechCrunch AIMEDIA · 2026-10-06TechCrunch quotes Mistral VP Pierre Stock saying ML4 used 'only 4,000 NVIDIA GPUs,' a different figure than 3,800.
CONTEXT The DecoderMEDIA · 2026-10-06The Decoder confirms training occurred in European data centers but does not provide a definitive GPU count, so it provides contextual confirmation of location rather than the exact GPU number.
SOURCES & TIMELINE
3Today, we’re launching a public preview of Mistral Large 4. Unofficially ML4, very officially: le Chonk . ML4 pushes the frontier of open-weight performance. You can try the preview API today on Mistral Studio . Weights drop end of this month. ML4 is a 1 trillion-parameter natively multimodal model with 49 billion active parameters. It is our largest and most capable model to date, and it continues to improve rapidl…
The race between open and closed AI is on, and Europe is still in the mix. On Tuesday, French AI lab Mistral AI has released Mistral Large 4 , a new large multimodal model aiming to leapfrog both American and Chinese rivals — following what French president Macron has described as “a third way in AI.” Amid a growing divide between closed models that can be unplugged and open models that are often made in China, Mist…
Mistral has released a preview of Mistral Large 4, a one-trillion-parameter model trained in Europe. Weights are expected at the end of October. In the Artificial Analysis Intelligence Index, the model jumps to 38 points but still trails leading closed models like Claude Opus 5.5 by a wide margin. Mistral's main selling point is IT security: the model reproduces and patches software vulnerabilities that competitors…