Knowledge catalogue
Model management is the real SDLC scaling bottleneck. @FactoryAI standardized on Fireworks to solve it: → 2–3x open-model growth → 5–15x more work per dollar → day-0 access to every new open-weight mo
Fireworks AI is now live on EvoSkill v1.3.0! You can now use @FireworksAI_HQ directly with EvoSkill to run fast inference on open models as both the evolution harness backend and the LLM scorer. Along
The big lesson from training @cursor_ai Composer 2: models exploit flaws in their training environment before learning what you actually want. Real RL for coding agents means production-faithful envir
We hosted the first RSI RL Environments hackathon with @hud_evals @ycombinator and it was a blast! We watched builders treat RL as a general purpose tool and reach for it across domains we had not ant
Well said @RamaswmySridhar! I believe the cost saving is actually much bigger, more like 4-5x. E.g., we have just reduced our GLM 5.2 cached token price by 2X to return efficiency gain to our users. A
This post compares the cost efficiency of Fireworks' GLM-5.2 model versus Anthropic's Opus across cached sessions, showing GLM-5.2 achieving approximately 4x lower cost per session at 90% cache hit ra
Fireworks AI announced the release of an open source GLM model that offers significant improvements in speed, cost efficiency, and output quality compared to existing alternatives. The post suggests t
In a joint Fireworks and @Faros_AI evaluation of 211 real engineering tasks, Claude Code + GLM-5.2 beat both Claude Code + Opus 4.8 and Codex + GPT-5.5: - Judge score: 0.568 vs. 0.521 and 0.466 - Time
Open + closed models = better together. Our previous research with @harvey showed the benefits of combining a frontier closed model as an advisor agent with fine-tuned, open-source worker agents. Thre
RL fine-tuning is now live for @nvidiaai Nemotron 3 on Fireworks, starting with Nemotron 3 Super (LoRA). Train with GRPO and serve the model in one place. We price by GPU-hour, not per token, so long
The underrated part of this announcement is that Fireworks has been quietly great behind the scenes helping us eval and serve these models Big kudos to the Fireworks team! Kimi K2.7 Code and GLM 5.2 a
GLM 5.2 is the open coding model everyone's been talking about. Now you can fine-tune it on Fireworks. SFT, DPO, and RL all supported. A leaderboard winner can still lose on your codebase. Training cl
It's way easier to switch models than to switch harnesses, and like many of you we use @cursor_ai every day. Now you can try out the latest open-source frontier model without changing your workflow. Y
Dr. Oz discusses the rapid pace of AI development and the challenge of staying informed about advances in the field, noting that even he uses AI tools to help manage and synthesize the growing body of
The hard part of reinforcement learning on a frontier model is the infrastructure that keeps training and inference numerically identical: zero KLD, end to end. We've solved this challenge, and are no
Bring open models to where you're already working. FireConnect brings Fireworks' top models directly into Claude Code, Pi, OpenCode, and Codex. Watch our Head of AI Education @Prof_oz show you how: Ho
GLM-5.2 has been the most popular new model on Fireworks this past week. @ArtificialAnlys confirms why: #3 overall on GDPval-AA (1524 Elo), #1 open weights by 116 points. Interest is showing no signs
You don't need to be AI Dependent to build State of the Art software. @FactoryAI and Fireworks use GLM to build the most advanced agentic AI you can train and own today. GLM 5.2 is available in Droid,
A user shares positive early impressions of GLM (likely a language model), praising its solid performance and ease of setup on Fireworks AI's platform. The post highlights Fireworks AI's developer exp
Fireworks AI reports that GLM 5.2, their language model, demonstrates performance comparable to GPT 5.5 across various evaluation benchmarks and real-world use cases. This claim suggests GLM 5.2 is co
Coding agents break when models are 'almost' bug-free. But almost valid JSON is just not the same valid JSON. Fun piece here from @akshay_pachaar shows why SFT can't fix this, and how GRPO trains agai
Fireworks Training Platform keeps expanding. Leading US open weight model Nemotron 3 Ultra is now ready for post-training: SFT and DPO via LoRA or full-parameter, on the same infrastructure that serve
Fireworks was named to @Redpoint's InfraRed 100 which recognizes the companies building the foundation for the next wave of AI. We're just getting started. Come build with us: https://fireworks.ai/car
Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 198B sparse MoE VLM designed by @StepFun_ai for inference from the start. 196B language backbone with a 1.8B v
NVIDIA Nemotron 3 Ultra is on Fireworks, day zero. Nemotron Ultra is an open model for frontier reasoning and orchestration in long-running autonomous agents. Think use cases like coding agents, deep
We spent the week at #MSBuild talking about one thing: fine-tuning has gone from 'maybe not worth it' to your actual competitive moat. @lqiao sat down with @yina_arenas to break down why and what Fire
Day 2 at #MSBuild is about what it takes to move beyond generic foundation models. Think customization, inference performance, and getting production-ready AI deployed at scale. @chahvivi will lead a
Fine-tuning to production inference is the gap where teams get stuck. At #MSBuild today, our own Rob Ferguson, @danielhanchen (@UnslothAI) and @marksaroufim (@coreautoai) discuss: model customization
Frontier models are powerful advisors. On @harvey's Legal Agent Benchmark, a GLM 5.1 worker using Claude Opus 4.7 as a sparse advisor reached 18/100 all-pass versus 14/100 for Opus alone, at 39% of th
MiniMax M3 arrives with MiniMax Sparse Attention (MSA), 15.6x faster decoding at 1M tokens. We're partnering with @MiniMax_AI to power the inference behind this week's launch. Head to http://minimax.i
Working with @FireworksAI_HQ to make MAI models easy to fine-tune and fully yours. Microsoft MAI models. Coming soon to Fireworks. Intelligence you control. End-to-end lineage you can prove. Fine-tune
Microsoft MAI models. Coming soon to Fireworks. Intelligence you control. End-to-end lineage you can prove. Fine-tune MAI reasoning models for your enterprise tasks. Your data. Your custom models. You
Move from test to production by running high-performance inference directly on Foundry. At #MSBuild, we demoed an end-to-end workflow showing how unified infrastructure improves latency, reduces cost,
Super excited to announce seven new world-class MAI models today. They represent what we consider a new era in AI designed to keep you in control and on the frontier. First is our text foundation mode
We’re looking forward to seeing how developers and enterprises use Fireworks AI on @Microsoft Foundry to power the next generation of intelligent applications. Catch us at booth F111 at #MSBuild and s
Many research labs only consider inference efficiency after the fact. Step 3.7 Flash is a 196B MoE model, and built for inference from the start by @StepFun_ai. Multi-Matrix Factorization Attention (M
Production AI systems require fundamentally different infrastructure approaches as workloads scale, with demands that differ significantly from development or testing environments. Fireworks AI discus
Another proof point for the open-weights thesis. From @RampLabs: 'If we built this again, we'd lean more on open-weight models.' Ramp pointed 10K agents at their own backend. Kimi K2.6 and DeepSeek V4
Reliability shouldn't require reserving GPUs. Serverless 2.0 is live on Fireworks: one API, 3 serving paths. → Standard: elastic default → Priority: sheds last under congestion, pricing ~1.5x standard
This tracks. 30 trillion tokens a day on our end, and open model share keeps climbing. Our partners @FactoryAI are seeing what we're seeing. Narrative violation: Open model use in Factory has more tha
10/ The bigger point: your product is the best RL environment you'll ever have. Frontier labs ship models that are good at everything. The opportunity is a model that's great at your thing. Product, u
9/ Real-time RL is where it gets fun. Catch live signals from real users on real generations. Update continuously. Ship a new version every few hours. Only works if the base model is already good enou
Fireworks is coming to Tech Week, and we're doing it across multiple cities. Boston is up first. We're co-hosting a Founder Skybar Social with @fin_ai and @TrustVanta. Follow along via #BOSTechWeek In
Cursor's new Composer 2.5 takes third on the Artificial Analysis Coding Agent Index and is ~10-60x lower cost than the higher-effort Opus 4.7 and GPT-5.5 variants above it. This release puts Composer
Fine-tuning used to mean a team, a GPU cluster, and weeks of iteration. Now it's just a CLI command, ~10 min of GPU time, a few cents of compute. You walk away owning the weights. Open models off the
Nathan's @cursor_ai team didn't prompt-engineer their way to Composer 2.5. They trained it. The massive RL program runs RL rollouts on Fireworks, alongside production inference. 'Comment 🔥 to see my p
Fireworks is coming to Tech Week First up: Boston. We're co-hosting a rooftop founder social with @fin_ai and @TrustVanta — curated crowd, real conversations, limited space. Thu May 28 · 7–10pm · #BOS
We ran 720 browser agent tasks with @nottecore across frontier models. One baseline model produced malformed outputs in ~1 out of every 5 calls, leading to retries inside multi-step workflows. Across
Not all the good stuff at Build happens on the main stage. Join Microsoft for Startups on June 2 for Dev Your Own Way. Hands-on activations with @GitHub, @FireworksAI_HQ and @NVIDIAforStartups. Real c
The @cursor_ai team shipped Composer 2 and now Composer 2.5 on the same Kimi K2.5 base model. Performance benchmarks are📈. Frontier quality and open-source economics. 85% of the compute powering these
a new era of hackathons: 2005→ look what i built with web search 2016 → can you build a rails app in 24 hrs? 2025 → spin up a CX bot with Claude in 5 min 2026 → train your own model over a weekend hap
Fireworks AI posted about a billboard at their headquarters, likely showcasing their company branding, a product announcement, or marketing campaign. The post appears to be promotional content shared
free to start. fast on day zero. proud to power the default model for LangSmith Fleet. happy building. LangSmith Fleet now has a free model powered by @FireworksAI_HQ for Developer and Plus plans. It’
Weekends are for vibe coding. But are your vibes continuously improving? Fine-tune your own model → stop waiting on someone else's release cycle. Today's training update: Gemma 4 Dense is now availabl
Fireworks Training Platform continues to expand. Today GLM 5.1 LoRA RL is now live via Training API: SFT, DPO, and full RL on a 200K context window → custom loss functions or smart defaults. No usage
If you're calling a third-party API, your competitor can make the same call tomorrow. @lqiao is the keynote speaker at @pycon tomorrow, breaking down how to build a durable moat using fine-tuned model
Kimi K2.6 and DeepSeek V4 Pro are now GA on @FireworksAI_HQ on Foundry + PTU support in the US Data Zone—predictable performance, data residency, enterprise SLAs. Frontier open models. Azure controls.
Like any AI dev not employed by a closed lab, we share the ambition that at least 10x more should be capable of training frontier models in 2026. And like we all know, some 10x leaps can't survive a c
Most teams can pick frontier models. Fewer can run them at production scale without hitting constraints in latency, throughput, and governance. Fireworks AI on @Azure AI Foundry provides the inference
𝐅𝐮𝐥𝐥-𝐏𝐚𝐫𝐚𝐦 𝐑𝐋 𝐧𝐨𝐰 𝐚𝐯𝐚𝐢𝐥𝐚𝐛𝐥𝐞 𝐟𝐨𝐫 𝐊𝐢𝐦𝐢 𝐊𝟐.𝟔 You've been told only 3 AI labs matter. The best AI apps never believed that. @cursor_ai, @vercel, @genspark_ai don't run only off-the-shelf models. They trai