November 2026 Summaries

Fireworks updated its October 9, 2026 security bulletin after an October 23 incident at a third-party cloud provider enabled unauthorized access to Fireworks development machines and an unauthorized workload that affected additional customer data. Working with Palo Alto Networks Unit 42 and law enforcement, the company linked the activity to a financially motivated group associated with “LLMJacking,” in which stolen access tokens are sold to illicit LLM routing services. Beyond previously identified stored secrets, the investigation found that affected internal systems exposed resource metadata, account and billing information, and, for a limited number of customers, secrets, datasets, and incomplete model weights; payment card numbers were not stored or affected. Fireworks said its Inference, Training SDK, and Nexus services were not impacted and use zero data retention by default, while it has notified affected customers, revoked and replaced credentials, restricted network pathways, strengthened data-access controls and monitoring, and temporarily disabled self-service dataset uploads and downloads for Batch Inference and dataset-based training.

Moonshot AI’s Kimi K3.1 Pro is now available through Fireworks as a pay-per-token serverless model or via reserved-capacity dedicated deployments. Built from Kimi K3 and updated with sparse-attention technology, the text-and-vision model has 2.78 trillion total parameters, 104 billion active parameters, and a 1.04-million-token context window designed to improve long-context serving efficiency. Fireworks reports that it occupies the quality-for-cost frontier across five Specialized Intelligence Index benchmarks, ranking first in banking customer support, top among open models for legacy code, incident response, and corporate legal work, and tying Claude Opus 5.5 in finance. Compared with Kimi K3, it showed gains across all five listed benchmarks, including a 13.2-point improvement on Sierra’s banking evaluation. Moonshot also reports strong results in long-horizon coding, tool-assisted research, and office-work benchmarks, though these comparisons rely partly on published competitor scores and specific evaluation harnesses. Standard Fireworks serverless pricing is $2 per million uncached input tokens, $0.20 for cached input, and $10 for output tokens.

Fireworks announced several products at its first Forge conference aimed at helping organizations build, customize, and deploy AI using open-weight models and their own data. Fireworks Ultra-fast, powered by NVIDIA Blackwell GPUs and in public preview, is designed to accelerate latency-sensitive agent workloads, while Background Mode, multi-day prompt caching, and higher rate limits seek to reduce inference costs and support greater concurrency. The company also introduced Nexus Desktop and expanded Nexus integration with tools including Codex, Claude Desktop, and OpenCode, providing model routing, spending visibility, and enterprise controls. New training capabilities include No Eval, which converts production signals into task-specific evaluations, and Specialized Model Inference, which serves LoRA-adapted models through a serverless API. For enterprise customers, Fireworks made governance, spending controls, and audit logs generally available, previewed regional data and encryption features, and launched Fireworks Sovereign in private preview to enable managed inference and training within customer-controlled cloud or on-premises infrastructure.