The latest AI news and updates of Fireworks AI — AI inference platform for fast, customizable model serving and compound AI systems at scale. Covering Fireworks AI's latest product updates, launches, and analysis from the past 90 days.
Fireworks AIOct 24Fireworks published research on optimizing reinforcement learning for Moonshot AI's 2.8T-parameter Kimi K3, achieving up to 2.61× faster routed-expert forward passes and ~99% prompt-token cache reuse. The platform now offers serverless training for Kimi K3 through its Training API, supporting task and reward function testing on shared capacity without dedicated GPU provisioning.
Fireworks AIOct 23Fireworks reports that Heidi post-trained NVIDIA’s Nemotron 3 Ultra on its platform, achieving a 54% win rate against the frontier model currently powering Heidi Evidence in clinician side-by-side evaluations. The 12-person team used synthetic data for supervised fine-tuning and reinforcement learning. The quantized model, optimized for NVIDIA Blackwell, is currently in final evaluation ahead of a production rollout.
Fireworks AIOct 21Fireworks AI fine-tuned Meta's Muse Glimmer 30B on Kalshi market outcomes, outperforming Jev by 11.2 percentage points in prediction accuracy on unseen topics. In a hypothetical backtest, the tuned model returned +7.1% while Jev lost 8.1%. Fireworks now offers free supervised fine-tuning for Muse Glimmer 30B through November 30 as part of its new Free Weights Club.
Fireworks AIOct 10Fireworks AI identified a security incident involving unauthorized use of internal credentials that exposed environment variables, including API keys and tokens, in evaluation jobs for a limited number of customer accounts. The company found no evidence of access to models, inference traffic, or personal data. Affected customers must rotate identified credentials and check for unauthorized activity.
Fireworks AIOct 5Fireworks AI has added DeepSeek V4.1 Flash to its Dedicated Training API and Managed Training surfaces. The model serves as a base for agentic coding, terminal automation, and tool use. Fireworks notes that the model is cost-efficient to serve in production environments.
Fireworks AISep 30Fireworks AI added Z.ai's GLM 5.3 Flash to its Serverless Training API, supporting vision and text fine-tuning on a shared GPU pool. The model performs well on Fireworks' benchmarks for agentic coding, document analysis, and tool use. The platform provides immediate access to the model for training without the need for dedicated infrastructure provisioning.
Fireworks AISep 29Fireworks AI launched FireRouter with Opus, a cache-aware router that directs coding tasks between Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash. The system achieves 98.1% of Opus accuracy while reducing cost per coding session by 57%. It is available as a standalone serverless endpoint and through the Fireworks CLI for integration into existing coding harnesses.
Fireworks AISep 24Fireworks Research released Ember-1, a specialized model built on Kimi K3 that delivers top-tier quality with roughly 40% fewer tokens. By shortening reasoning traces, the model reduces costs for agentic coding and reasoning workloads. Ember-1 is available today as a Research Preview on the Fireworks Serverless platform, with training support for custom, token-efficient model development.
Fireworks AISep 24Fireworks released ARCv3, a compressor that reduces BF16 weight update sizes by nearly 50% compared to its previous version. This lossless compression keeps rollout fleets synchronized with the trainer's current policy, making cross-region reinforcement learning more practical. The update is available in the Fireworks Training API via the fireworks-delta-compression package for teams using their own trainer.
Fireworks AISep 22Fireworks AI launched the Specialized Intelligence Index, a platform aggregating real-world, domain-specific benchmarks authored by industry practitioners. The index covers sectors including security, legal, finance, and healthcare, featuring contributions from organizations like Doximity, Harvey, and Sierra. It provides comparative data on model quality, cost, and task duration to help teams select models based on production-grade performance.
Fireworks AISep 22Fireworks AI analyzed 18 models across 113 coding tasks on the DeepSWE benchmark. The analysis found that routing each task to the model that handled it best achieved a 97.6% solve rate at $1.88 per task, compared to 74.1% at $6.52 for the best single model. This hindsight result demonstrates that capability is split across the existing model pool.
Fireworks AISep 17Fireworks AI added Z.ai's GLM 5.3 to its Serverless Training API, supporting LoRA fine-tuning on a shared GPU pool without provisioning. The model joins Kimi K3 and Qwen 3.8 27b, with GLM 5.3 Flash coming soon. The platform provides immediate access through the Fireworks cookbook and quickstart documentation, requiring no sales demos or DMs.
Fireworks AISep 15Fireworks AI released DeepSeek-V4.1-Flash, a 552B-parameter MoE model that matches GPT-6 Astra-level coding accuracy on the DeepSWE benchmark at 1/15th the cost per task. The model features a split encoder-decoder architecture and improved KV cache utilization, targeting the high input-token and cache-hit costs typical of long-running agentic coding workflows.
Fireworks AISep 15Fireworks AI reports that Juicebox reduced latency by 80% and annual inference costs from $4M to $800K. The deployment utilized specialized models tailored to Juicebox’s use case, which involves running multiple real-time search agents across hundreds of thousands of talent profiles.
Fireworks AISep 14Fireworks AI now hosts Alibaba's Qwen 3.8 Max, a mixture-of-experts model built on the Qwen 3.5 foundation. The model supports function calling and image input, available through serverless API and on-demand deployment. Pricing starts at $2.00 per million input tokens and $6.00 per million output tokens.
Fireworks AISep 10Fireworks AI now hosts DeepSeek V4.1 Flash, a 552B-parameter multimodal mixture-of-experts model. It features a 1-million-token context window and a causal encoder-decoder architecture for coding, cybersecurity, and agentic tasks. Fireworks reports the model outperforms Opus 5 and GPT 5.6 Sol on key benchmarks at one-fortieth the cost, with dedicated training support arriving soon.
Fireworks AISep 10Fireworks and Genspark released Gen-1 Slides, an open model post-trained from a MiniMax M3 base using long-horizon reinforcement learning. The model matches Claude Opus 5 performance on slide generation at approximately 1/17 of the input-token price. It is now the default model in Genspark AI Slides, delivering improved visual design and reduced layout defects in production.
Fireworks AISep 8Fireworks AI reports that partner Cosine AI’s Lumen Outpost coding model, trained on Fireworks Training, outperforms GPT-5.5 by over 3x on cost per successful task. The model specializes in niche languages like Fortran and Verilog where generic models struggle. Cosine AI uses the Fireworks platform to run its entire training and inference pipeline end to end.
Fireworks AISep 1Fireworks and LangChain Labs fine-tuned a Qwen-3.5-35B model to detect perceived errors in production agent traces. The specialized judge matches the performance of frontier models like GPT-5.5 and Claude Opus while operating at up to 100x lower cost. The model uses managed supervised fine-tuning on Fireworks infrastructure to identify user corrections and rejections within multi-turn agent interactions.
Fireworks AISep 1Heidi Health fine-tuned open models on Fireworks AI to power its ambient clinical scribe, achieving higher clinician preference than Gemini. The specialized models reduced transcription latency from 25 seconds to 7 seconds. This deployment utilizes the Fireworks Training API, which is now generally available for organizations to build and serve custom models on frontier-grade infrastructure.
Fireworks AIAug 31FactoryAI used the Fireworks AI Training API to fine-tune an open Qwen model for its Droid Shield secret detection tool. This specialized model caught nearly 20% more real secrets than GPT-5.5 while reducing cost and latency. The Training API is now generally available for organizations to build and serve custom models on frontier-grade infrastructure.
Fireworks AIAug 31Fireworks AI has moved its Training API to general availability and launched Fireworks Lab, a service for co-designing and building specialized models. The platform integrates the serving stack into the training loop to manage rollout throughput and weight synchronization, enabling organizations to train models that outperform frontier systems on domain-specific tasks.
Fireworks AIAug 29Fireworks AI has released Z.ai's GLM-5.3-Flash for general availability on its inference platform. The company delayed the launch by two days to investigate an unexplained reasoning-length discrepancy, prioritizing output quality over speed. The model features 320B total parameters, a 1040k-token context window, and supports native multimodality, including image input and function calling.
Fireworks AIAug 29Fireworks AI launched day-zero support for Z.ai's GLM 5.3 on its US-hosted serverless inference platform. The model delivers a 50% improvement over GLM 5.2 on Z.ai's Code Bench and provides state-of-the-art open-weight performance for cybersecurity and complex coding tasks. It also achieves top-tier scores on the Artificial Analysis Agentic Index.
Fireworks AIAug 26Fireworks AI and Harvey report that Tenet, a legal-specialized model post-trained on a Kimi K3 base, achieved a 19.7% all-pass rate on the Legal Agent Benchmark, up from 10.8%. The model reached state-of-the-art performance on LAB Contracts while maintaining a flat cost per task. These gains generalize to benchmarks not seen during training, including Apex Agents and Redline Bench.
Fireworks AIAug 26Fireworks AI reports that DeepSeek V4 Pro outperforms Fable 5 on SWE-Bench and LiveCodeBench while reducing costs per solved task by two-thirds. The model also completed 840 adversarial security runs with zero refusals or output truncations. DeepSeek V4 Pro is available now on Fireworks for serverless inference and dedicated training.
Fireworks AIAug 22Fireworks AI now supports Muse Glimmer 30B on its Dedicated Training API, enabling both LoRA and full-parameter fine-tuning. The platform provides reserved GPU capacity for the U.S.-developed, open-weight model, which is optimized for agentic workflows, reliable tool use, and long-horizon reasoning tasks.
Fireworks AIAug 14Fireworks AI launched the official DeepSeek-V4-Pro-0813 model, featuring a 1.6-trillion-parameter mixture-of-experts architecture and a 1.04-million-token context window. The model includes a speculative decoding module and outperforms Claude Opus 4.8 on agentic benchmarks like Terminal Bench 2.1 and DeepSWE at 6x the cost efficiency. It is available now via serverless API and on-demand deployment.
Fireworks AIAug 13Fireworks AI is partnering with Arcee.ai to bring the 2.8-trillion-parameter Kimi-K3 model to Arcee’s new open models API beta. This integration supports Arcee’s newly open-sourced nac agent harness, which manages complex, multi-step engineering workloads. The partnership provides access to Kimi-K3’s native vision capabilities and 1-million-token context window for autonomous agent tasks.
Fireworks AIAug 13Fireworks AI applied Anthropic’s J-Lens probe to Kimi K3 and Qwen 3.5-9B, revealing that these models internally prepare relevant vocabulary before generating text. This research confirms that open-weights models exhibit measurable silent signals mid-draft, mirroring internal reasoning patterns previously observed in proprietary systems. The findings demonstrate that internal model states can be decoded to visualize concepts forming before output.
Fireworks AIAug 12Fireworks AI launched day-0 support for Alibaba’s Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter sparse mixture-of-experts model. The model features a 262k-token context window and is optimized for autonomous agents and complex coding tasks. It is available via serverless API and on-demand deployment, with serverless pricing set at $2.00 per million input tokens and $6.00 per million output tokens.
Fireworks AIAug 10Fireworks AI now hosts the full Voyage AI model lineup, including the Voyage 4 family, voyage-multimodal-3.5, and rerank-2.5. This integration enables a complete retrieval-to-generation pipeline on a single platform. By consolidating embedding, retrieval, reranking, and generation, the platform reduces latency and simplifies data boundaries for teams building AI on proprietary data.
Fireworks AIAug 10Fireworks AI launched a training skill that integrates model fine-tuning directly into AI coding agents. The skill allows Claude Code, Cursor, and Codex to configure, validate, and launch training jobs from plain-language goals. It provides cost estimates and troubleshooting support within the agent environment, requiring only the Fireworks CLI and an API key for setup.
Fireworks AIAug 4Fireworks AI evaluated Kimi K3 on the CyberGym-E2E security benchmark, finding it outperformed frontier models by avoiding refusal and policy blocks. Initial low scores were traced to poor harness integration; after enabling native tool calling and preserving reasoning state, Kimi K3’s S1 success rate rose from 40.7% to 70.0% and S4 patch generation reached 54.5%.
Fireworks AIJul 31Fireworks AI conducted controlled experiments on Qwen3.5-9B to compare LoRA and full-parameter fine-tuning. The study identifies data coverage, learning-rate optimization, and adapter rank as key levers for closing quality gaps. Fireworks AI suggests evaluating outputs and tuning training recipes before increasing adapter rank, as these interventions often resolve performance differences without the higher cost of full-parameter fine-tuning.
Fireworks AIJul 31Fireworks AI launched a contrastive fine-tuning recipe for the Qwen3-Embedding-8B model, allowing adaptation to specialized domains for under $10. The platform uses in-batch negatives to train models on private query-positive pairs, delivering performance gains such as a 36% nDCG lift on legal citation retrieval tasks.
Frequently asked questions