The Pulse
Dream Releases MiST Cybersecurity Models With 13.1-Point Gain
Dream released two open-weight MiST cybersecurity models, including an 8B model that scored 61.7 against 48.6 for its Qwen3 baseline. The company says the 8B model’s average rose 13.1 points across 13 cybersecurity tasks, while its 30B-A3B
AI.info Team · Written with AI · fact-checked ·
The final MiST checkpoints improve mean cybersecurity accuracy by +13.1 and +8.6 absolute percentage points over the corresponding Qwen baselines for 8B and 32B, respectively, corresponding to relative gains of +27.0% and +15.8%.
Oded Ovadia, Elad Ben Zaken, Elad Guttman and Orly Moreno Kadosh, authors of MiST: Mid-Training LLMs for Cybersecurity
Two MiST models are available
Dream announced the MiST family on October 19, releasing MiST-8B and MiST-30B-A3B as downloadable model weights. The release accompanies the research paper “MiST: Mid-Training LLMs for Cybersecurity,” which describes an 8B and a 32B Qwen-based model; the 30B-A3B release instead adapts NVIDIA’s Nemotron 3.5 Lightning checkpoint. The 8B model card identifies its released checkpoint as MiST-8B-DPO, while the 30B-A3B card identifies its checkpoint as MiST-30B-A3B-SFT.
Dream’s headline 13.1-point result belongs to the 8B model, not the 30B-A3B release. On the company’s 13-task cybersecurity suite, MiST-8B-DPO averages 61.7, compared with 48.6 for Qwen3-8B, its base-model comparison. The paper reports the change as a 13.1 percentage-point gain, or 27% relative to the baseline.
Scores span six cybersecurity benchmarks
The evaluation covers 13 tasks drawn from CTI-Bench, SECURE, AthenaBench, PrimeVul, MMLU-Cyber and SecEval. Dream reports the paper’s results as averages over five runs, with reasoning modes disabled where available. Its table gives an unweighted mean across the tasks, rather than a result on one standalone benchmark.
The paper’s final 32B checkpoint averages 63.2, up from 54.6 for Qwen3-32B, an 8.6-point increase. That result is part of the research comparison; Dream’s announcement names the 8B and Nemotron-based 30B-A3B checkpoints as the releases. On a separate comparison reported with the 30B-A3B model card, that model scores 60.3 against 51.5 for its Nemotron base, an 8.8-point difference. The card says its results and the base model’s results average three runs.
Synthetic data supplies the specialist training
MiST’s method inserts a cybersecurity mid-training stage between general pretraining and later instruction tuning. Dream starts with a curated set of security materials, including vulnerability records and taxonomies, then expands that source material into synthetic explanations, question-and-answer examples and analytical exercises. The resulting mid-training corpus contains about 976 million tokens, according to the paper.
After mid-training, the team applies supervised fine-tuning, or SFT, to teach instruction-following and chat behavior. The paper says nearly all of the cybersecurity improvement is already present after mid-training and SFT: MiST-8B-SFT averages 62.4, slightly above the final DPO checkpoint’s 61.7. Direct preference optimization uses general preference data rather than cybersecurity-specific preferences, so the authors do not identify it as the main source of the domain gains.
The authors also compare MiST with a raw continued-pretraining baseline built from a larger cybersecurity text corpus. Their 8B MiST approach uses fewer training tokens yet performs better on nearly all the benchmarks in that comparison. The results point to the composition and transformation of domain material—not simply the amount of raw text—as a central part of the method.
Benchmark gains do not establish operational safety
Dream frames MiST for cybersecurity research, analysis and further adaptation, not as an autonomous security agent. The 30B-A3B model card says agentic workflows and live security operations were not specifically evaluated, and cautions that the results measure performance on benchmarks rather than readiness for deployment in operational settings. It also advises human oversight and additional safeguards for systems built on the model.
The paper includes a preliminary safety evaluation of the 8B checkpoints on the CyberSecEval MITRE benchmark. It reports safe-response rates of 89.8% for MiST-8B-SFT and 85.3% for MiST-8B-DPO, compared with 75.8% for Qwen3-8B. The authors describe that test as preliminary; it does not establish how the models behave under adaptive, multi-turn or tool-assisted attacks.
MiST-8B is listed under the Apache 2.0 license, and the 30B-A3B model card also lists Apache 2.0 while noting that inherited NVIDIA model materials remain subject to the separate OpenMDW-1.1 license. The concrete offer is therefore two downloadable starting points with different base models and training stages, accompanied by benchmark results—not a validated replacement for security operations or human review.