Brief

Fireworks AI has identified critical numerical parity bugs that can arise when training and serving large language models, particularly Mixture-of-Experts (MoE) architectures. These discrepancies, stemming from the non-associative nature of floating-point arithmetic and differing summation orders in distributed training versus inference, can lead to subtle but significant issues. Such drift can compromise the integrity of reinforcement learning from human feedback (RLHF) due to altered log probabilities and erode customer trust in fine-tuned models. AI

DeepSeek-V4 introduces novel training techniques, including Anticipatory Routing to stabilize training by using older weights for routing decisions, and a Generative Reward Model (GRM) where the model itself acts as a judge for complex tasks. The model also supports three distinct reasoning modes (Non-think, Think High, Think Max) trained with varied configurations for different reasoning depths. These advancements highlight the need for flexible, programmable training infrastructure that can adapt to complex, co-designed model and runtime systems. AI

Fireworks AI is expanding its presence by participating in Tech Week events across multiple cities, starting with Boston. The company is co-hosting a social event in Boston with Fin AI and Trust Vanta, inviting attendees to follow updates using the hashtag #BOSTechWeek. AI

LangChain has released updates for its Fireworks integration, with version 1.4.1 addressing API connection errors and retries. Version 1.4.0 introduced a migration to the 1.x SDK for Fireworks AI and included fixes for context overflow errors. These updates aim to improve the stability and reliability of using Fireworks models through the LangChain framework. AI

Fireworks AI has released Composer 2.5, an inference infrastructure update for its coding agent. This new version achieved third place on the Artificial Analysis Coding Agent Index. Composer 2.5 also offers a significant cost reduction, being 10-60 times cheaper than previous iterations. AI

Fireworks AI has introduced a significantly streamlined process for fine-tuning open-source models. What previously required substantial resources like dedicated GPU clusters and weeks of work can now be accomplished with a simple command in about ten minutes and minimal cost. This advancement aims to make custom model development more accessible, suggesting that readily available open models in 2026 will serve as effective starting points for various applications. AI

Fireworks AI is providing the inference infrastructure for Cursor AI's new Composer 2.5 model. Cursor AI's team trained the model using a large-scale reinforcement learning program that runs rollouts on Fireworks' platform. AI

Fireworks AI has released full-parameter reinforcement learning for Kimi K2.6, enabling custom model training. This move supports companies like Cursor, Vercel, and Genspark that train open-source models on proprietary data. The announcement highlights the growing trend of specialized AI applications moving beyond off-the-shelf solutions. AI

Fireworks AI, an inference infrastructure company, is participating in Microsoft's "Dev Your Own Way" event on June 2. This event is part of Microsoft's Build conference, highlighting that significant developments can occur beyond the main stage presentations. AI

Fireworks AI is sponsoring hackathons to encourage the development of AI applications. The company envisions a future where individuals can train their own AI models over a single weekend, building on the rapid progress seen in AI development from web search capabilities to complex bot creation. AI

Fireworks AI has launched a new inference infrastructure service designed for speed and cost-effectiveness. The service is free to start and aims to provide rapid performance from day one. It is already powering the default model for LangSmith Fleet, indicating early adoption by a notable platform. AI

Fireworks AI has announced that Kimi K2.6 and DeepSeek V4 Pro models are now generally available on its platform. These models are accessible via Azure Foundry and include PTU support within the US Data Zone, promising predictable performance for users. AI

Fireworks AI has launched its Training Platform, now supporting GLM 5.1 LoRA RL fine-tuning. The platform offers SFT, DPO, and full RL capabilities with a 200K context window. Users can leverage custom loss functions or default settings, with no usage limits or credit restrictions. AI

Fireworks AI is offering its inference infrastructure on Azure AI Foundry, aiming to help teams run frontier models at production scale. This solution addresses common constraints in latency, throughput, and governance that many organizations face when deploying advanced AI models. AI

Fireworks AI is partnering with LangChain to provide inference infrastructure for advanced agents. The collaboration was highlighted at the Interrupt 2026 conference in San Francisco. This partnership aims to support the development of sophisticated AI agents by ensuring robust underlying infrastructure. AI

Innovative Solutions, an AWS Premier Partner, has redesigned its enterprise services delivery by adopting Fireworks AI as its primary inference layer. This strategic shift addresses escalating AI inference costs and delivery complexity, which were previously limiting profit margins and operational flexibility. By moving its DarcyIQ platform to Fireworks AI, the company achieved predictable economics and enabled a transition from linear service models to parallel, agent-driven execution. AI

Fireworks AI has developed a new feature called 'safe_tokenization' to prevent prompt injection attacks in large language models. This technique ensures that user input, which can contain malicious control tokens, is treated as data rather than code by the model. By distinguishing between user-provided text and the model's internal control tokens, safe_tokenization maintains the integrity of prompt structures, preventing unauthorized alterations to model behavior. AI