September 2026 Summaries

Reinforcement learning at scale can become unstable when rollout generation and gradient training use separate implementations that differ in numerical precision, kernels, batching, parallelism, or operation order, causing them to assign different token probabilities despite identical model weights. Experiments cited with GLM 5.2 found that an unaligned trainer and inference engine produced a KL divergence of about 0.013, discarded roughly 45% of batches, and saw rewards collapse, while aligned calculations maintained zero KL and stable rewards. Mixture-of-Experts models add further risk because small rounding differences can alter expert routing and because even identical expert selections may yield divergent outputs if aggregation precision or execution order differs; a Qwen3.5-MoE test traced a substantial discrepancy to expert-output aggregation. Such failures may resemble issues with data, reward design, or learning rates, making probability-level validation important. The post argues that reliable RL infrastructure should co-develop training and rollout systems, align numerical operations, reuse routing decisions where needed, and measure full-model training–inference agreement through metrics such as KL divergence.

Fireworks Nexus has introduced FireRouter with Opus, a cache-aware model router designed to reduce AI coding costs while maintaining performance by selecting among Claude Opus 5.5, GLM 5.3, and GLM 5.3 Flash for each user turn. Available through the Fireworks CLI and as a standalone serverless endpoint, the router evaluates task suitability, model costs, and prompt-cache tradeoffs to balance quality and expense. In internal A/B testing on coding workloads, FireRouter reduced average session costs by 57%, from $15.36 to $6.63, while achieving 78.7% graded-turn accuracy compared with 80.2% for Opus-only usage, or 98.1% of Opus’s overall accuracy. Its routing produced a 94.2% cache hit rate versus 97.8% for Opus alone, representing a deliberate reduction in cache retention to gain broader cost savings. The service supports tools including Claude Code, Codex, and Cursor IDE, with its model selection expected to evolve as new models become available.

Fireworks has introduced a GLOBAL deployment option that allows inference workloads to use compatible GPU capacity across more than 30 regions and multiple cloud providers through a single deployment identity and endpoint, reducing the need for customers to manage separate regional deployments, routers, and failover systems. The system plans multi-region placement before requests arrive, while nearby healthy gateways use prepublished routing information to send traffic to available compute and retry eligible failures without adding a global scheduler step to each request, though physical network distance can still affect latency. Customers can select GLOBAL or narrower US, Europe, APAC, and Canada boundaries according to data residency, compliance, hardware, cost, and reliability requirements. Fireworks reports that, in a seven-day observational production analysis, globally placeable deployments operating across two or more serving clusters had a 99.992% request success rate compared with 99.269% for single-region deployments, while noting that the findings are not causal or a universal uptime guarantee. The redesign addresses a prior limitation in which deployments needed all requested nodes within one region, enabling workloads to combine capacity across regions when no single location has sufficient resources.

Fireworks announced ARCv3, a lossless compressor designed to reduce the BF16 model-weight updates transferred from reinforcement learning trainers to geographically distributed rollout machines. Based on 1,000 production RL update deltas, ARCv3 achieved an average compressed size of 0.19% of the original weights, compared with 0.36% for ARCv2, 0.35% for PULSESync, and 0.67% for a draft Hugging Face TRL implementation. The approach exploits the observation that only about 2% of weights change between checkpoints and that most changes affect mantissa bits rather than a weight’s exponent or sign, omitting unchanged values and separately encoding common and rare update types. By shrinking transfers while reconstructing weights bit-for-bit identically, the system is intended to help rollout fleets remain aligned with the latest policy across multiple regions using shared object storage and standard network links rather than a single co-located cluster. ARCv3 is the default for Fireworks Trainer SDK rollouts and is also available through the fireworks-delta-compression package for users operating their own trainers.

Fireworks Research introduced Ember-1, a specialized model derived from Kimi K3 that it says maintains comparable quality while using roughly 35–50% fewer reasoning tokens, targeting lower costs for coding and multi-turn agent workloads. The company trained the model through more than 50 experiments and 200 evaluations, aiming to preserve useful reflection and feedback-driven planning while removing excessive reasoning traces that can compound costs across agent interactions. Fireworks reports that Ember-1 matched or approached K3’s performance across public and industry benchmarks, including clinical, software engineering, terminal, and airline tasks, while often improving the quality-to-cost tradeoff relative to K3’s reasoning settings and several other proprietary models. Live A/B tests with two production coding customers reportedly showed about 35% lower tokens per task at comparable quality, and an internal rollout was unnoticed by developers despite reduced token use. Ember-1 is available as a Research Preview on Fireworks Serverless alongside Kimi K3, with planned training support for organizations seeking customized, token-efficient versions.

Fireworks AI has launched the Specialized Intelligence Index (SII), a platform comparing open, closed, and specialized AI models on practitioner-designed benchmarks covering healthcare, legal work, finance, cybersecurity, customer support, productivity, and software engineering. The index argues that conventional benchmarks often fail to reflect real workplace performance because they rely on bounded tasks and standardized scoring, whereas professional work involves incomplete information, changing requirements, judgment, multi-step processes, and operational constraints. SII therefore emphasizes evaluations built with domain practitioners, explicit acceptance standards, realistic tools and conditions, validation against expert review, and reporting of reliability, costs, duration, and critical failures. Rather than creating a single cross-domain ranking, it presents benchmark-level results and tradeoff views so organizations can assess models for their own deployment needs. Benchmark owners can contribute evaluations in a compatible format, while Fireworks reviews and runs submitted benchmarks across qualifying production-scale models before publishing approved scores.

Fireworks argues that the effective frontier for coding agents may come from routing tasks among complementary models rather than relying on a single strongest model. Its analysis of 18 models on the 113-task DeepSWE v1.1 benchmark found that an hindsight-based “oracle” router could achieve a 97.6% success rate at an estimated $1.88 per task, compared with GPT-6 Astra’s 74.1% at $6.52, although the company notes that this oracle is upwardly biased because it selects winners after observing multiple outcomes. The analysis suggests that expensive models are uniquely best for relatively few tasks, while a curated portfolio of only a few models captures much of the potential gain, with the best three reaching 91.2% in the oracle evaluation. It also emphasizes that real-world routing is difficult because a router must predict the appropriate model before execution, and unreliable selection can be worse than consistently using one model. Fireworks presents FireRouter as a task-level, cache-aware routing system for open and closed models, reporting that its internal production coding traffic cost 53% less over four weeks than using Claude Opus 5 alone.

Phylo, an applied research lab behind Biomni Lab, uses long-horizon AI agents to help biologists plan, execute, and document research involving large datasets, specialized tools, HPC resources, and hundreds of database integrations. Originally reliant on proprietary models, the company adopted a model-agnostic approach that evaluates open and closed frontier models for quality, latency, and cost, routing much of its workload to open-weight models served through Fireworks. Phylo reports that Fireworks’ serverless endpoints, rapid access to newly released models, and engineering support enabled it to move models into production within about 24 hours, reduce costs by 60%, roughly halve time to first token, and support 2x month-on-month user growth. While proprietary models remain preferred for the most difficult tasks, Phylo says open models perform well across much biomedical work and give users more choice without model lock-in. The company plans to use millions of real-world agent traces to fine-tune and train specialized open models, while advising AI founders to focus engineering effort on customer problems, agent workflows, and product integrations rather than operating their own inference infrastructure.

Fireworks reports that DeepSeek-V4.1-Flash establishes a strong cost-performance position for autonomous software engineering, achieving a 74.34% DeepSWE pass rate comparable to GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 while costing about $0.43 per task, or roughly 15 times less than Astra. The 552-billion-parameter mixture-of-experts model uses a split activation design with 8 billion active input parameters and 16 billion output parameters, aiming to reduce costs for coding agents that repeatedly consume far more input context than they generate. Improved KV-cache efficiency, including lower HBM and SSD requirements, is presented as especially important because cached input represented most spending in long agent trajectories. On Terminal-Bench 2.1, it reportedly reached 86.5% accuracy, one percentage point below Astra, at about one-twelfth the total cost per task. However, it performed less strongly on Humanity’s Last Exam, scoring 34.52% compared with Astra’s 50.40%, suggesting it is better suited to agentic coding than difficult academic reasoning. An oracle-routing evaluation found that combining both models could reach 54.80% on HLE, indicating that their differing strengths may improve results in multi-model systems.

Specialized intelligence is presented as a progression from using rented closed frontier models to building and training models tailored to a company’s unique data, workflows, taxonomy, style, and operational needs. Closed APIs offer powerful initial capabilities but can create dependencies around cost, latency, data control, pricing, and model changes, while prompt engineering, retrieval-augmented generation, tool use, and agent harnesses can improve performance without changing the underlying model. Teams may then incorporate open models to gain greater control and improve unit economics, using task-specific evaluations rather than general benchmarks to select or route work among models based on their relative strengths. Training or fine-tuning becomes valuable when organizations need models to learn proprietary classifications, writing conventions, efficient behaviors, or domain-specific knowledge, while distillation can transfer large-model performance into smaller, cheaper models and reinforcement learning can improve agents on verifiable multistep tasks. For domains that change frequently, the central advantage of ownership is a continual training loop that incorporates new data, production failures, and stronger base models, allowing specialized intelligence to evolve rather than depend on external providers’ roadmaps.

Genspark and Fireworks Lab post-trained the open-weight MiniMax M3 model into Gen-1 Slides, a specialized agentic system designed to plan, create, render, review, and revise presentation decks across long, multi-turn workflows. The partners treated slide generation as a reinforcement learning challenge because high-quality decks require visual judgment, self-correction, and credit assignment across trajectories that can exceed 100,000 tokens, rather than simple imitation of finished examples. Genspark defined production-based quality standards and evaluation criteria covering design, layout failures, and factual issues, while Fireworks Lab developed training infrastructure, reward engineering, and stability measures through more than 100 experiments. Training used a staged curriculum that began with supervised fine-tuning on curated decks, then expanded RL context lengths as the model improved, while addressing numerical problems such as tokenization mismatches between inference and training systems that could bias updates. According to Genspark’s internal evaluation, Gen-1 Slides matches or exceeds Opus 5 on most quality measures while costing roughly one-seventeenth as much per input token and about 90% less per completed deck, reducing low-rated production decks from 18% to 3.6% relative to the base model.