TAU-HOME.COM
LOADING

Fireworks AI Unveils 1,500 TPS NVIDIA Blackwell Inference Engine and Free 30B Model Fine-Tuning

At its inaugural Forge conference, Fireworks AI introduced its NVIDIA Blackwell-powered Ultra-fast engine in public preview (up to 1,500 TPS), half-price Backgr

tau · November 13, 2026

#Fireworks AI #Blackwell #LLM Inference #Fine-Tuning #DeepSeek #AI Agents

Fireworks AI Unveils 1,500 TPS NVIDIA Blackwell Inference Engine and Free 30B Model Fine-Tuning

On November 13, 2026, opening day of its first annual Fireworks Forge conference, Fireworks AI unveiled a suite of inference, training, and enterprise control products designed to help organizations build specialized AI on open-weight models. Headline releases include the public preview of Fireworks Ultra-fast on NVIDIA Blackwell GPUs alongside half-price asynchronous execution and time-limited free fine-tuning.

High-speed LLM inference architecture on NVIDIA Blackwell GPUs and agent workload routing diagram unveiled at Fireworks Forge

Image source: https://fireworks.ai/blog/fireworks-unveils-new-products-at-forge-to-help-companies-build-specialized-intelligence

When AI agents make dozens of sequential model calls to finish a single workflow, latency and token expenses compound quickly. Fireworks AI structured its Forge announcements around separating time-critical interactive calls from deferrable background workloads.

NVIDIA Blackwell Ultra-Fast Engine and Workload-Matched Pricing

The flagship hardware-backed release is Fireworks Ultra-fast, powered by NVIDIA Blackwell GPUs and now available in public preview. Fireworks AI demonstrated generation speeds of up to 1,500 tokens per second (TPS) on DeepSeek V4.1 Flash. In an internal benchmarked 30-turn agent task, end-to-end completion time dropped from 1 minute to 8 seconds using Ultra-fast.

For workloads where lower unit cost outweighs immediate turnaround, Fireworks introduced new runtime and caching options:

  • Background Mode: Applies a 50% price discount to eligible deferrable workloads, such as generating evaluation datasets or running overnight research tasks.
  • Multi-day Prompt Caching and Throughput Scaling: Multi-day prompt caching reduces repeated context reprocessing costs for persistent agents, while higher serverless rate limits and Elastic Burst—which lets customers exceed normal usage ceilings when spare capacity is available without code or pricing changes—are now generally available.
  • Specialized Ember-1 Model: Available on serverless and dedicated deployments, Fireworks' Ember-1 model used roughly 40% fewer tokens than Kimi K3 at comparable quality in company evaluations, and scored within about one point of Claude Opus 5 across five industry benchmarks at roughly one-fifth the cost per task.

Pairing Ultra-fast for live multi-turn agents with half-price Background Mode for batch jobs allows engineering teams to allocate inference spend based on task urgency rather than paying peak interactive rates across every pipeline.

Free Muse Glimmer 30B Fine-Tuning Through November 30 and Serverless LoRA

Fireworks AI also rolled out new tooling to train and serve task-specific models without provisioning dedicated GPU clusters. Through November 30, 2026, teams can run free fine-tuning on the Muse Glimmer 30B model covering supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning (RL).

Serving those customized weights moves into public preview alongside an automated evaluation tool:

  • Specialized Model Inference (Public Preview): Extends serverless per-token inference to LoRA adapters, starting with GLM 5.3 Flash. Teams incur zero serving charges while an adapter sits idle and can keep a production adapter live while testing newer iterations.
  • No Eval (Private Preview): Converts production execution traces and product signals into task-specific evaluations so teams can pinpoint failure modes and measure post-training gains without spending weeks manually labeling data.

Nexus Desktop Launch and Enterprise Control Plane

To bring model routing directly into daily employee workflows, Fireworks launched Nexus Desktop alongside the existing Nexus CLI. Nexus Desktop connects the Codex app, Claude Desktop, and OpenCode with a single toggle per tool, combining open and closed models with centralized spending visibility. Pro plans start at $8 per month, Teams plans start at $10 per seat per month, and Enterprise plans add SSO, governance, shared budgets, and access controls.

On the governance front, model governance, spending controls, and exportable audit logs are now generally available, while centralized account management, regional data controls, and customer-managed encryption are in preview. For regulated organizations requiring on-premises or customer-hosted execution, Fireworks Sovereign entered private preview, bringing the Fireworks inference and training stack onto customer-owned cloud, neocloud, or data center GPUs; Fireworks noted it already manages more than 25,000 GPUs inside customer environments.

Verified Scope and Preview Caveats

Several core components from the Forge announcement remain in preview or carry explicit eligibility limits. Both Fireworks Ultra-fast and Specialized Model Inference are in public preview, while No Eval and Fireworks Sovereign are restricted to private preview. Free SFT, DPO, and RL training on Muse Glimmer 30B runs strictly through November 30, 2026, and the 50% Background Mode discount applies only to eligible asynchronous workloads that do not require immediate completion.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285