Is MiST state of the art?Pending benchmarks.Announced Oct 19, 2026; no independent results yet.

The details so far

Lab

Dream

Announced

Oct 19, 2026

API

Not yet

Open weights

Published

Parameters

8B, and 30B total with about 3B active (MoE)

Models

MiST-8B (built on Qwen3-8B-Base) and MiST-30B-A3B (built on NVIDIA Nemotron 3.5 Lightning 30B-A3B Base)

Training

Cybersecurity mid-training on about 976M synthetic tokens, then SFT; the 8B adds a DPO step

Focus

Threat intelligence, CVE-to-CWE mapping, MITRE ATT&CK and vulnerability analysis

Status

Research models, separate from Dream's production models; not for autonomous agent use

Paper

MiST: Mid-Training LLMs for Cybersecurity (EMNLP 2026)

License

Apache-2.0 (the 30B-A3B's NVIDIA base materials stay under OpenMDW-1.1)

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab-reported scores (self-reported): MiST-8B averages 61.7 across 13 tasks from six public cybersecurity benchmarks, vs 48.6 for Qwen3-8B; MiST-30B-A3B averages 60.3, vs 51.5 for its Nemotron base.

Lab claim: MiST-8B beat a 70B cybersecurity model on Dream's 13-task evaluation.

Sources