News Hub

Fireworks AI offers Kimi K3.1 Pro with improved cost efficiency and speed.

Fireworks AI unveiled new products at their Forge event.

Numeric alignment and mixture-of-experts routing matter for reinforcement learning alignment.

Fireworks AI introduced FireRouter with Opus.

Fireworks AI introduced Ember-1.

ARCv3 demonstrates cross-region reinforcement learning for efficient model optimization.

Fireworks AI released Specialized Intelligence Index benchmarking real-world AI workloads.

Fireworks AI evaluated model routing across 18 models on DeepSWE benchmark.

Fireworks AI compared DeepSeek-V4.1-Flash and GPT-6 Astra on coding benchmarks.

Gen-1 Slides generates Opus 5-level presentation decks at reduced cost.

Fireworks AI released Training API supporting custom reinforcement learning and fine-tuning loops.

DeepSeek V4 Pro and Fable 5 compared on SWE-Bench software engineering tasks and cost per task metrics.

DeepSeek V4 Pro evaluated on CyberGym security agent tasks with cost measurements reported.

Fireworks AI post-trained Kimi K3 model with Harvey for long-horizon legal work applications.

Researchers reproduced J-Lens readouts on Kimi K3 and Qwen3.5-9B to detect silent signals before model output.

Fireworks AI offers performance stack combining Fireworks Models with Voyage AI embeddings.

Three tests to run before switching from LoRA to full fine-tuning.

Guide for fine-tuning embedding models to improve domain-specific retrieval performance.

Kimi K3 offers US-only serverless inference and day-zero training capabilities.

Trilogy released Kimi K3, an open-weight cybersecurity model.