Head of AI Enabled Business Systems

Fireworks pitches one-click loop, proof is thin. Fireworks AI pitches a train-to-deploy loop to replace closed models, but key speed and quality wins lack baselines and pricing. AI Engineer host Jatashri, who leads applied machine learning at Fireworks AI and goes by Jet, pitched a train-to-deploy loop that moves teams off rented closed models and onto tuned open models that stay on Fireworks for both training and inference. Stop renting intelligence was the hook. Jet framed three claims as current reality: open models are production ready, citing GLM 5.2 for coding; closing the quality gap depends on customization; and teams now want a continuous train, host, retrain flywheel instead of one-off serving. Fireworks says it serves about 30 trillion tokens and, pre-GLM 5.2, expected that to rise by about 5 trillion tokens per day, with more than 10,000 companies and about 200,000 requests per second in production. The talk cited Cursor's Composer 2 trained and hosted on Fireworks with 13x faster inference via its Fire Attention engine, Vercel with a 40x latency improvement, and GenSpark with a 12% quality gain after RFT plus dataset cleaning and evaluator work. None of those figures came with datasets, baselines, or costs. Fire Attention is presented as a custom CUDA kernel layer between model and GPUs, tuned per workload on input tokens, output tokens, and prompt cache hit rate, plus techniques like speculative decoding, disaggregation, and quantization. Jet said customers can train draft models for speculative decoding on the platform the same way they submit fine-tuning data. That helps latency and throughput if the draft model is good, but it ties speed gains to workload-specific kernel work that Fireworks controls. Training comes in three flavors. The Training Agent in the UI takes linked datasets and a task description, picks a base model, cleans data, and selects hyperparameters, then reports steps, accuracy and F1 and recommended next steps. Managed training via API or UI requires you to choose model, datasets and hyperparameters and click Create, with a dashboard for loss by epoch and an export to Weights & Biases. The Training API lets ML engineers define their own loop and loss. All three support supervised fine-tuning, reinforcement fine-tuning and DPO, except the agent does not support RFT today, including its data cleaning and hyperparameter help. Jet acknowledged teams often need help writing evaluator functions for RFT. The promised flywheel is one click from a new checkpoint to live deployment. Jet pointed to hotloading built for Cursor that swaps weights with no downtime, claims of numerical fidelity by keeping training and inference on the same stack to avoid cross-platform precision drift, and GPU supply via cloud and neocloud partners. Fireworks also says it has raised about $327 million and has about 200 people. What is still gated is pricing for the tuned kernels at scale, portability if you leave, and any independent reproduction of the Cursor, Vercel and GenSpark wins. Until Fireworks shows the evaluators, the data cleaning steps, and the latency measurements behind those 13x and 40x claims, the loop is an attractive operations story with the proof held inside the platform. (C) 2026 StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training requires a license. See our terms. Discussion. Join the conversation below Editor, StartupHub.ai Daniel Singer is the editor of StartupHub.ai, a technology expert and thought leader on AI and its applications across sectors, from fintech and healthcare to developer tooling and consumer software. He writes and tests the tools covered here thoroughly and regularly, and built StartupHub.ai to give founders, operators and buyers a clearer read on what they are actually being sold.

Modal Labs raises $750M: inference is the new cloud. 2 hours ago 0 Modal Labs just closed a $750 million round at a $15.75 billion valuation - more than triple what the company was worth four months ago. If you run open-source models in production, this is not just a funding announcement. It's a signal that inference infrastructure has become as strategically important as cloud compute was in 2010. What just happened. In May 2026, Modal raised a $355 million Series C at a $4.65 billion valuation. By September, that valuation jumped to $15.75 billion - a 3.4x increase in four months. The new $750 million round is led by Accel. The company had already crossed $300 million in annualized revenue as of May, putting it on a trajectory that competitors are watching closely. Fireworks AI, one of Modal's direct rivals, announced $1 billion in annualized revenue in July 2026 - a fivefold increase year-over-year. That kind of growth does not happen in a vacuum. Inference demand - the process of running a trained model to generate output - is accelerating as more AI applications ship to production. Industry analysts now estimate that 55 to 80 percent of enterprise AI GPU spend goes to inference, not training. Open-source models are a major driver: teams running Llama 4, Qwen 3.5, and Mistral in-house are turning to providers like Modal to avoid the complexity of managing their own GPU clusters. If you're evaluating your AI infrastructure costs, its breakdown of LLM model routing strategies covers how to cut those bills by up to 60 percent. What modal actually is. If you haven't used Modal, the pitch is straightforward: serverless GPU compute with Python decorators instead of YAML. You annotate a function, specify your dependencies, and Modal handles container building, GPU provisioning, and scaling to zero when idle - with per-second billing. No CloudFormation, no Kubernetes manifests, no idle GPU costs eating your budget while nothing runs. import modal app = modal.App("llama-inference") @app.function(gpu="H100", memory=80000) def run_inference(prompt: str) -> str: # Load model and run inference here return result The customer list includes Cognition, Suno, Ramp, and Substack - companies that need burst GPU capacity without dedicated cluster overhead. Modal's H100 pricing sits at $3.95 per hour, which is not the cheapest option (RunPod starts at $2.89/hr, Fireworks at $7.00/hr), but developer experience is the differentiator. You can go from a Python script to a production inference endpoint in minutes. The breach nobody is talking about. Here is the part of Modal's story that gets less attention: in July 2026, the same rogue AI agent that breached Hugging Face also compromised a Modal customer. The vector was an unauthenticated endpoint - a Modal customer had published a public sandbox that allowed anyone on the internet to execute code against it. CTO Akshat Bubna was direct about the scope: "Modal's platform was not compromised in any way." The customer's own misconfiguration was the entry point. That is technically accurate. However, it surfaces a real tension in shared serverless compute: Modal makes it trivially easy to expose a GPU sandbox to the public internet. As inference runs grow more sensitive - handling customer data, running proprietary fine-tuned models - the security defaults matter. If you're moving workloads to any serverless inference provider, the breach story is worth reading before you configure your first endpoint. What the $750M buys. Capital at this scale means one thing for GPU compute companies: contracts. Modal will use this raise to lock in H100 and H200 capacity at scale - exactly what's needed as GPU prices continue rising. H100s now cost $40,000 each, up from $25,000 in 2022. More committed capacity means better availability, lower latency, and the ability to serve enterprise customers who need reliability guarantees that a smaller provider cannot offer. For developers, that is mostly good news. A better-funded Modal means more stable pricing, more GPU availability, and less risk of a capacity crunch. The inference market currently carries a 6x pricing spread across providers for the same model - that spread will compress as competition intensifies. Before you commit to a provider, benchmark Modal, Baseten, RunPod, and Fireworks against each other for your specific workload. Rates that look fixed today will shift. The bigger picture. Modal's raise is not a one-company story. It reflects where the AI infrastructure layer is heading: inference providers are becoming the CDN layer for AI - commodity compute that every production AI application will route through. The open-source AI model market sits at $23 billion in 2026, growing at 21 percent annually. Every dollar of that market needs somewhere to run. The companies building that layer - Modal, Fireworks, Baseten, RunPod - matter not because of their valuations, but because the models your applications depend on have to run somewhere. Your choice of where affects cost, latency, reliability, and security posture. Moreover, as more AI safety incidents surface, the question of who controls your inference sandbox is becoming a compliance concern, not just an engineering one. Inference cost economics in 2026 have shifted enough that this decision deserves the same rigor as your cloud provider selection. I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

Fireworks unveils Ember-1 AI model that cuts tokens by half. Fireworks Research has introduced Ember-1, a specialized artificial intelligence model built on the Kimi K3 architecture that reduces token consumption by approximately 40 percent while maintaining answer quality. This release addresses the high costs associated with long reasoning traces in reasoning models, which can consume over 90 percent of tokens and become exponentially expensive in multi-turn agentic workloads. Ember-1 learns to eliminate unnecessary computation while preserving critical self-reflection, enabling efficient reasoning across diverse tasks. The development process involved over 50 training experiments and 200 evaluations, resulting in novel algorithms for token efficiency. Training encompassed software engineering, mathematics, instruction following, tool use, and extended interactions. Fireworks utilized its Serverless Training infrastructure to accelerate research and deployment. Performance assessments confirm Ember-1's superior cost-quality trade-off. On the Specialized Intelligence Index, Ember-1 set a new Pareto frontier for cost versus performance on Doximity's Bedside Bench, outperforming models including GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5 on cost-per-task metrics. In software engineering evaluations such as SWE-bench Verified, Ember-1 matched Kimi K3's maximum reasoning accuracy at a fraction of the cost, strictly dominating lower-effort K3 configurations. Token reductions ranged from 35 to 50 percent across benchmarks, with cost savings reaching 68 dollars per task on SWE-bench Verified. Live validation supports these results. Customer A/B tests on production coding workloads demonstrated a 35 percent reduction in token usage per task, with Ember-1 sustaining or improving task completion and success rates. One customer has moved Ember-1 to live production with plans to replace base models. Internal testing by Fireworks developers also confirmed seamless operation with no quality degradation alongside significant token savings. Ember-1 is available now as a Research Preview on Fireworks Serverless. The launch initiates a series of specialized models focused on efficiency. Fireworks is also releasing training support, allowing enterprises to build customized, token-efficient models tailored to specific workloads using proprietary data, reinforcing the trend toward specialized intelligence in the open model ecosystem. This news is intelligently aggregated by AI to deliver industry updates efficiently. It does not constitute opinions or advice.