Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- Fireworks unveils Ultra-fast inference, Background Mode, and Nexus Desktop at Forge
At its first Forge conference, Fireworks announced products aimed at specialized open-model deployments: Ultra-fast serverless inference in public preview on NVIDIA Blackwell (up to about 1,500 tokens/sec on DeepSeek V4.1 Flash, with a cited 30-turn agent task falling from about 1 minute to 8 seconds), Background Mode at half price for deferrable work, multi-day prompt caching, higher rate limits, and Fireworks Nexus Desktop alongside the Nexus CLI so open and closed models can run inside tools such as the Codex app, Claude Desktop, and OpenCode with spend visibility. Pro plans start at $8/month and Teams at $10/seat; Enterprise adds SSO, governance, shared budgets, and access controls. The post also cites Ember-1 token-efficiency results versus Kimi K3.
- Fireworks serves DeepSeek V4.1 Flash 1017 at more than 1,500 tokens per second on its Ultrafast tier
Fireworks says DeepSeek V4.1 Flash 1017 runs at over 1,521 tokens per second on Fireworks Ultrafast, which is open by waitlist.
- Fireworks AI will retire GPT OSS 120B, Inkling and two Nemotron models from Serverless on November 7
Fireworks AI says several Serverless models will be decommissioned on November 7, 2026: GPT OSS 120B, Inkling, NVIDIA Nemotron 3 Ultra, NVIDIA Nemotron 3.5 Lightning 30B A3B, and the US endpoints for MiniMax M3 and GLM 5.3 Flash. It recommends dedicated deployments or replacements such as Muse Spark Open, Nemotron Super VL and DeepSeek V4.1 Flash 1017; dedicated deployments and Reserved Throughput are unaffected. Fireworks also lists DeepSeek V4.1 Flash 1017 and Muse Spark Open as new on Serverless and says MiniMax M3.1 is coming after its open-source release, expected Friday, October 23.
- Fireworks AI raises its Small serverless rate-limit tier to models under 800B parameters, moving GLM 5.3 into it
Fireworks AI's Small serverless model tier now covers models with fewer than 800B total parameters, up from 600B, per its docs changelog dated October 18. The Medium tier now spans 800B to under 1.6T parameters, which moves models such as GLM 5.3 into the Small tier. The rate-limit ceilings for each tier are unchanged; Fireworks doubled the Small tier's ceilings on October 7.
- Fireworks AI adds NVIDIA Nemotron 3.5 Super VL as a training model with up to 262K context
Fireworks AI says NVIDIA Nemotron 3.5 Super VL 120B A12B is now available for training on all three of its surfaces, with up to 262K context: Managed Training (LoRA SFT and DPO), the Dedicated Training API (LoRA and full-parameter training, including RL) and Serverless Training (LoRA on the shared pool, billed per token), per its docs changelog.
- Fireworks Forge session: Armadin post-trains open models to monitor security agents
Fireworks AI previewed a Nov 3 San Francisco Forge talk by Armadin Security’s Yonatan Oren on post-training open models to monitor agents that hold production permissions.
- Engram Lab co-founder to speak on agents with native memory at Fireworks Forge on Nov 3
Fireworks AI says Engram Lab co-founder @realJessyLin will speak at its Fireworks Forge event in San Francisco on November 3 about building agents that remember things in their weights as well as their context.
- Interconnects author @natolambert to speak at Fireworks Forge on how open models win
Fireworks AI says @natolambert, co-founder of Trillium Labs and author of Interconnects, will speak at its Forge event about why so many teams are building on open models.
- Fireworks congratulates Zeta Global on launching Athena Inference Model on its stack
Fireworks AI congratulated Zeta Global on launching the Athena Inference Model (AIM), citing it as another specialized model served to customers on Fireworks.
- Fireworks introduces FireRouter with Opus, its first router model
Fireworks AI introduced FireRouter with Opus, its first router model: routine tasks go to top open models while Claude Opus handles the rest.