Fireworks

Just another example of how companies are bringing specialized intelligence to their customers with the help of Fireworks. Congrats on the launch of Athena Inference Model (AIM) @ZetaGlobal !

Just another example of how companies are brining specialized intelligence to their customers with the help of Fireworks. Congrats on the launch of Athena Inference Model (AIM) @ZetaGlobal !

Every serious AI team eventually builds their own data, evals, models, and maybe even chips. @swyx of @latentspacepod and @aidotengineer on why the future is domain-specific everything. Join us Nov 3 in SF: bit.ly/3TxgUuB

Yes, it’s true! @dhh is speaking at Fireworks Forge sharing his new thoughts on AI coding. In a fireside chat, we'll talk controlling the AI stack and what that means for small teams building specialized intelligence. Join us Nov 3 in SF: bit.ly/3TxgUuB

New models and agents ship every week. How do you know which ones actually work for your product? @sarahmsachs, Head of AI at @NotionHQ, makes the case for building your own evals at Forge. Nov 3, SF. Apply to attend: bit.ly/3TxgUuB

Turns out you don't need a frontier lab budget to build a solid decision classifier. Our Head of AI Developer Education @prof_oz fine-tuned a 9B Jev-style classifier on Fireworks using plain SFT, public datasets, and two simple tricks. The full recipe is open source, so you can train your own.

New in Nexus: Savings See how much Fireworks open models and FireRouter are saving your team, by user and model. Plus, track your cache hit rate. Check out Nexus: bit.ly/4y3vrfW

DeepSeek V4.1 Flash is now available for training on both the Dedicated Training API and Managed Training surfaces! It's a strong base for agentic coding, terminal automation, and tool use. It's also very cost-efficient to serve. Get started today: docs.fireworks.ai/fine-tuning/mo…

Rollouts drive most of RL's compute cost. But splitting rollout and training across engines risks numerical mismatches, and in MoE models, that can even send tokens to different experts. We co-build both engines at Fireworks, so training stays fast and consistent. Learn more: t.co/k9z3pSUh3I

Rollouts drive most of RL's compute cost. But splitting rollout and training across engines risks numerical mismatches, and in MoE models, that can even send tokens to different experts. We co-build both engines at Fireworks, so training stays fast and consistent. Learn more: t.co/YVqOShLAU5

Congrats to the @Fulcrum_inc team on Echo! Final training runs were done entirely via Fireworks' serverless training API, on a Kimi K3 LoRA (rank 32). Make sure to read their writeup, which includes the "persona post-training" approach and how they landed a $4.5k total training cost.

GLM 5.3 Flash is now available for training on the Serverless Training API - accessible to all, with both vision and text support. This model performs great on our benchmarks for agentic coding, document analysis, and tool use while being cost-efficient to serve. Get started today: t.co/CUCSSE8Dnz

Independent numbers from @ValsAI. Fewer tokens per turn means lower cost and faster agent loops on every call. Try Ember-1 on Fireworks Serverless: fireworks.ai/models/firewor…

Discover when to leverage the GLOBAL option and how we architected multi-region routing to handle global workload placement off the inference hot path, optimizing for low latency, resilience and operational ease. x.com/i/article/2105…

Fireworks is at @AIconference! Find us at Booth #245 (Shed A) and swing by our Coffee Lounge in the Innovation Hub. Catch our own Jetashree Ravi's breakout session, and @RobFergusonIII on the Day Two keynote stage. See you on the show floor!

The Fireworks Forge speaker lineup continues to grow! Join us in welcoming @NotionHQ's Head of AI, @sarahmsachs, creator of Ruby on Rails and CTO of 37signals, @dhh, and the Founder of Instinct, @noahrshinn! Check out our other announced speakers here: bit.ly/3TxgUuB

Attendance is free, but space is limited. Apply to attend: bit.ly/3TxgUuB

Introducing FireRouter with Opus. Our first router model. Routine tasks go to top open models, Claude Opus handles the rest. Cache and task-aware routing. Proven 98.1% of Opus accuracy with 57% lower cost per coding session. Read more: fireworks.ai/blog/introduci…

SF, come out tomorrow night for rooftop drinks and open office hours with the Fireworks team. Want to walk out having actually fine-tuned, evaluated, and deployed a model on Fireworks that night? We can do that too. Bring your laptop. RSVP: luma.com/270tizh3

Huge congrats to @noahrshinn on Instinct's raise announcement! What incredible momentum for the team. Catch Noah live as he takes the stage at Fireworks Forge, you won't want to miss it. Apply to attend here: fireworks.ai/forge#apply

Today, @normalfactoryco joins the Specialized Intelligence Index with CAD Arena. Can AI turn an engineering drawing into a CAD part that's accurate and editable? Normal tests agents across five CAD platforms, expanding the SII into engineering design.

See how the most popular models stack in up your domain. Explore the SII: fireworks.ai/specialized-in…

MiMo-V2.6-Pro is available today as an on-demand deployment. → Highest-scoring open-weight model on the AA Intelligence Index (46) → 1.02T total / 42B active MoE → 1M context, multi-modal, with MIT license Spin up an endpoint and tell us what you’re building → t.co/4Df7Foijnt

Want to RL-train a model on your own task without building the infra? Use our new cookbook with @hud_evals: → Define your task + grader in HUD → Sample + train the model on Fireworks Define the task once. Train and evaluate against the same environment. t.co/Bdi3A2XWjr

Proud to be a launch training partner for @LangChain's LangSmith Fine-Tuning and smithtune. LangSmith traces → managed SFT on Fireworks → eval → deploy. One CLI, no training infra. Get a Fireworks API key: fireworks.ai/api-keys Then run smithtune: github.com/langchain-ai/s…

Today, @FactoryAI joins the Specialized Intelligence Index with Legacy-Bench! The SII keeps growing, now with benchmarks testing how models handle the legacy software behind payroll, financial settlement, and more. Submissions are open for teams with proprietary benchmarks. Get listed: t.co/r4eBbGb8Hu

This just in: @JeffTangney, CEO and Co-founder of @doximity is joining the Fireworks Forge speaker lineup! Join us on Nov. 3 in San Francisco for a day with the builders, researchers, and technical leaders pushing AI forward. The event is free to attend, but space is limited.

Apply to attend: fireworks.ai/forge?utm_sour…

Ember-1 is a specialized model from Fireworks Research designed to make every token go further. Built on Kimi K3, it produces shorter reasoning traces, using roughly 40% fewer tokens while maintaining top-tier quality. fireworks.ai/models/firewor…

Read more about Ember-1 here: fireworks.ai/blog/ember-1?u…

Training can cut costs by up to 90% on the work your customers depend on. The next model drop won’t get you there, no matter how great it is. Join us and @Macroscope's @Rob_Bishop as we chat about owning the intelligence at the core of your product. fireworks.ai/event/build-li…

ARCv3 is Fireworks' new compressor for RL. It cuts the size of BF16 weight updates by nearly 50%, which makes cross-region training more practical and keeps rollouts current. For teams using their own trainer with Fireworks rollouts: fireworks.ai/blog/arcv3-and…

The speaker lineup at Fireworks Forge keeps getting stronger. Welcome @pirroh, President & Head of AI @Replit, to an already stacked roster! Join the people, teams, and companies building their own frontier on open models. Nov. 3, San Francisco.

Apply to attend: fireworks.ai/forge#apply

Today we're launching the Specialized Intelligence Index (SII): one destination for real-work benchmarks across industries, built by the teams that use them every day. Hear from Fireworks co-founder @the_bunny_chen on the importance of specialized benchmarks:

Vídeo

Featured on SII: @doximity @mercor @harvey, @depthfirstlabs, @novee_security, @genspark_ai, @SierraPlatform, @DecagonAI, @ProximalHQ, @Macroscope, @RogoAI, @traversal_ai fireworks.ai/index

We ran 18 models across 113 real coding tasks on DeepSWE, then went back and asked a simple question: what if every task had routed to the model that handled it best? Answer: 97.6% solve rate at $1.88 per task, versus the best model at 74.1% at $6.52. The next frontier is a router. Full analysis: t.co/yzpA9I3V39

Don’t miss newly announced Fireworks Forge speaker: @BrendanFoody, CEO and Co-founder of @Mercor. Join the people, teams, and companies building their own frontier on open models. Nov 3, San Francisco. Apply to attend: fireworks.ai/forge#apply

Behind every production agent is a serious stack. Join @e2b, Fireworks, and @braintrust on 9/30 for technical talks on sandboxing untrusted code, serving models at agent-loop speed, and knowing whether a change actually helped. RSVP: luma.com/e2b-0e34

We’ve just added three new speakers to our incredible Fireworks Forge lineup! We are proud to welcome @BrendanFoody @pirroh and @jefftangney to the stage as we bring together the people, teams, and companies building their own frontier on open models.

Apply to attend: fireworks.ai/forge#apply

Open models are becoming increasingly capable in biology. Serving them well is harder. Fireworks gives @phylo_bio day-zero access and a fast serverless path, putting research agents in front of more scientists. Learn more: fireworks.ai/blog/phylo-bri…

Come hang out with AI engineers, tech leads, and the Fireworks team during AI Engineer 2026. Want to leave with hands-on experience fine-tuning, evaluating, and deploying a model on Fireworks? Bring your laptop. We'll make it happen. Join us: luma.com/gydwrqt6

GLM 5.3 is now available for training on the Serverless Training API, joining Kimi K3, Qwen 3.8 27b and more. GLM 5.3 Flash coming soon. No sales demo or DMs required. Here to help as needed, but you can also spend that time building. Start with docs or a pre-made recipe ➔ t.co/URmbsVlNp2

Quickstart docs: docs.fireworks.ai/fine-tuning/tr…

Many assume that agent spend goes toward output tokens. When we ran DeepSWE on Astra vs. DeepSeek V4.1-Flash, input tokens outnumbered output 174 to 1. 99.6% were cache hits. Those hits are 60% of the bill. Net result? Same quality. $0.43/task vs $6.52. x.com/i/article/2099…

Speed is key for @juicebox_work. They need to search hundreds of thousands of talent profiles in seconds with multiple real-time search agents. We helped them cut latency 80% and drop inference costs from $4M to $800K/yr, using specialized models tailored to their use case.

Vídeo

Excited to continue our partnership with the @alibaba_cloud team on this one, bringing the latest version of Qwen's flaship model to the Fireworks community. Start building with Qwen3p8 today: fireworks.ai/models/firewor…

Reinforcement learning at @cognition's scale is a hard infrastructure problem. We are proud to be part of the stack behind it. Congrats to the team on SWE-2! Read more about how we think about RL at Fireworks: fireworks.ai/blog/frontier-…

Recently, we hosted our 12th Nerd Meetup, this time at @lightfld 's office in SF. Builders came out from Lightfield, @qdrant_engine, and Fireworks. Building in the AI/infra space and want to join us for the next meetup? Follow along. We'll announce the next one soon!

GLM-5.3 is now available for training on Fireworks Dedicated Training API and Managed Training surfaces. Built for complex coding and long-horizon agents and now available to train on our dedicated infrastructure. Get started today: docs.fireworks.ai/fine-tuning/tr…

GLM-5.3 is now available for training on Fireworks' Dedicated Training API. Built for complex coding and long-horizon agents and now available to train on our dedicated infrastructure. Get started today: docs.fireworks.ai/fine-tuning/tr…

Deepseek-V4.1-Flash is available now on Fireworks! It is a 552B Parameter MoE built for coding, cybersecurity, and agents. It is the ideal workhorse model that outperforms Opus 5 and GPT 5.6 Sol at 1/40th the cost on DeepSWE,CyberGym, and Automation Bench! Interested in higher quality? Reach out, as we’re bringing this model to Fireworks Training soon. Try Deepseek-V4.1-Flash now on Fireworks: t.co/2Bez044BvU

Deepseek-V4.1-Flash is available now on Fireworks! It is a 552B Parameter MoE built for coding, cybersecurity, and agents. It is the ideal workhorse model that outperforms Opus 5 and GPT 5.6 Sol at 1/40th the cost on DeepSWE,CyberGym, and Automation Bench! Interested in higher quality? Reach out, as we’re bringing this model to Fireworks Training soon. Try Deepseek-V4.1-Flash now on Fireworks: t.co/2Bez044BvU

Introducing Gen-1 Slides: an open model that matches Claude Opus 5 on slide generation, at ~1/17 of its input-token price. @genspark_ai post-trained it from a @MiniMax_AI M3 base with Fireworks Lab using long-horizon RL. Live today as the default in Genspark AI Slides. 🧵

Genspark owned the product judgment. They brought their real production environment into training and shaped what the model optimized for: the standard for a good deck, the design principles behind it, and how to sharpen that standard over time.

Today, 10am PDT: we'll cover the journey from renting closed frontier models to owning your specialized intelligence, with a practical framework for knowing when to make the leap. Includes open Q&A w/ Head of AI Developer Education @Prof_OZ Register: fireworks.ai/event/webinar-…

Renting a closed model + a brittle, overfit harness only gets you so far. Routing to open models helps quality, cost, and latency, but you hit a ceiling. Most ambitious AI teams train their taxonomy, style, and judgment into models they own. Read more: fireworks.ai/blog/making-th…

We invite you to Forge. At Forge, we will challenge you to make your own frontier. You'll spend the day with the builders, researchers, and technical leaders pushing AI forward. Lin Qiao. Jensen Huang. Jay Parikh. And more. Apply to attend: fireworks.ai/forge

We invite you to Forge. At Forge, we will challenge you to make your own frontier. You'll spend the day with the builders, researchers, and technical leaders pushing AI forward. Jensen Huang. Lin Qiao. Michele Catasta. And more. Apply to attend: fireworks.ai/forge

Congrats to one of our Kimi K3 Fire Pass Community Hackathon winners @sethsaler for his winning entry: Dayprint! Our Head of AI Developer Education, the esteemed @Prof_OZ, walks through Seth's entry with live commentary. #kimik3onfireworks

Vídeo

Our partners at @CosineAI train their own coding model on real production code, for languages like Fortran and Verilog that generic models fumble. On cost per successful task, their Lumen Outpost model beats GPT-5.5 by more than 3x. The whole pipeline runs on Fireworks Training. Build your own frontier: t.co/E2RBk9thSV

Our Head of AI Developer Education @Prof_OZ is back with a new series! He’ll cover the journey from renting closed frontier models to owning your specialized intelligence, with a practical framework for knowing when to make the leap. Register here: fireworks.ai/event/webinar-…

Our co-founder @the_bunny_chen sat down with @FirasSozan to talk about why strong open models increase compute demand, not the other way around. Watch the episode: youtu.be/qb58MI9i1nE?si…

Research doesn’t end in right or wrong, but in insights that lead to more questions. General-purpose models aren’t fit for research because they don’t know customers’ workflows. @quadrillionlabs trains large custom models on its proprietary data using the Training API so its agent fits how researchers actually work. Build your own frontier: t.co/E2RBk9sK3n

We're going to need a bigger park. Fireworks is hiring across all teams. Check out our careers page, and come shape the future of AI with us! See you soon: fireworks.ai/careers

A noisy code reviewer gets ignored. Flag too many non-issues and engineers tune it out. @Macroscope got the control they needed with the Fireworks Training API to train a model using their own advantages and loss to improve the signal-to-noise ratio. Build your own frontier: t.co/E2RBk9thSV

Our partners @LangChain generate billions of tokens of agent traces a day, their richest signal on how real users react. Judging every one with a frontier closed model like GPT-5.5 or Opus is too costly, so they fine-tuned a Qwen base model on Fireworks that matches it at up to 100x lower cost. Build your own frontier: t.co/E2RBk9sK3n

DeepSeek V4 Flash now comes with native vision on Fireworks. Available now on serverless and priority tiers. Better benchmarks than the text-only 0731 version at the same low price: $0.22 input / $0.66 output per 1M. Give your workhorse model eyes. Try it today → t.co/ubE2Z2AnZ5

DeepSeek V4 Flash now comes with native vision on Fireworks. Available now on serverless and priority tiers. Better benchmarks than the text-only 0731 version at the same low price: $0.22 input / $0.66 output per 1M. Give your workhorse model eyes. Try it today → t.co/ubE2Z2AnZ5

Train past the frontier. @tryheidi's ambient scribe drafts a clinical note the clinician has to trust enough to sign. They fine-tuned open models on Fireworks whose notes their clinicians preferred over Gemini’s in side-by-side reviews, at 3.5x lower latency. Build your own frontier: t.co/E2RBk9thSV

A secret that leaks into shipped code is a breach waiting to happen. @FactoryAI's droids write and commit code faster than any reviewer can keep up, so Droid Shield has to be trustworthy and low-friction: catch the real secrets, clear the false alarms. With the Training API, Factory fine-tuned an open Qwen model that caught almost 20% more real secrets than GPT-5.5 at lower cost and latency. Build your own frontier: t.co/E2RBk9thSV

All week we’re spotlighting teams training past the frontier on Fireworks. @Figma’s AI team is shipping features that change how millions of people design, and with the Training API they can iterate faster. Less time debugging infra, more time on research. Build your own frontier: t.co/E2RBk9thSV

The most ambitious companies are training models to outperform the frontier on the capabilities that differentiate their business. Today, we announce the general availability of our Training API and Fireworks Lab, making model specialization accessible to all organizations.

Working with ML teams around the world surfaced where training workflows break down: constrained model and method choice, limited loop control, idle and costly compute, and split training and serving infrastructure that destabilizes runs Those partnerships shaped our Training API.

GLM-5.3-Flash: GA on Fireworks. We held it two days over a reasoning-length gap we couldn't explain. Same weights ≠ same model. Quality first. Start building → fireworks.ai/models/firewor…

GLM 5.3 is now live on Fireworks, day zero: → 50% improvement over GLM 5.2 on t.co/wQkYTEKzIU’s Code Bench → State-of-the-art open-weight performance on cybersecurity tasks → Excels at complex coding and long-horizon tasks → US-Hosted Serverless on Fireworks Try it now: t.co/9RTwsylMMA

GLM 5.3 is now live on Fireworks, day zero → 50% improvement over GLM 5.2 on t.co/wQkYTEKzIU’s Code Bench → State-of-the-art open-weight performance on cybersecurity tasks → Excels at complex coding and long-horizon tasks → US-Hosted Serverless on Fireworks Try it now: t.co/9RTwsylMMA

Congratulations to @LotusDecoder for their winning submission to the Kimi K3 Community Hackathon. Their entry, MedRecap, gets the full @Prof_OZ hackathon judging treatment here:

Recently @Harvey introduced Tenet, its first model, trained for long horizon legal work. Harvey post-trained it from a Kimi K3 base in collaboration with Fireworks using async RL on our Training API. A thread on promising initial results for both performance and cost-efficiency🧵

On LAB, Harvey's Legal Agent Benchmark, agents produce finished legal deliverables graded against dozens of criteria. All-pass counts a task only if it clears them all. Tenet lifts all-pass from 10.8% to 19.7% over the Kimi K3 base, reaching state-of-the-art on LAB Contracts.

Recently @Harvey introduced Tenet, its first model, trained for long horizon legal work. Harvey post-trained it from a Kimi K3 base in collaboration with Fireworks using async RL on our Training API. Promising initial results for both performance and cost-efficiency:🧵

On LAB, Harvey's Legal Agent Benchmark, agents produce finished legal deliverables graded against dozens of criteria. All-pass counts a task only if it clears them all. Tenet lifts all-pass from 10.8% to 19.7% over the Kimi K3 base, reaching state-of-the-art on LAB Contracts.

Your closed model refused a task it should have done. Ours didn't. DeepSeek V4 Pro on Fireworks: beats Fable 5 on SWE-Bench and LiveCodeBench, a third of the cost per solved task, and it doesn't decline legitimate security-review work. Read more: fireworks.ai/blog/DeepSeekV…

Model reliability is the #1 reason agents fail to complete a task. Frontier models demonstrate capability, then can't deliver consistently at production scale. Sept 2, NYC: Learn to train a real model yourself, then upload it to @Azure Foundry. luma.com/fw.models.nyc