August 2026 Summaries

Fireworks AI announced the general availability of its Training API and Fireworks Lab, expanding tools for companies seeking to customize open models for specialized applications such as legal work, clinical documentation, coding, and security. The Training API lets ML teams run custom Python-based training loops while Fireworks manages distributed training, model rollouts, weight synchronization, deployment, and recovery, with serverless LoRA training billed per token and dedicated infrastructure for larger or full-parameter workloads billed per GPU hour. The company emphasizes reinforcement-learning alignment between training and inference, including consistent numerical formats, kernels, and mixture-of-experts routing, as well as asynchronous rollouts and compressed weight updates intended to improve GPU utilization and iteration speed. Fireworks also offers a unified train-deploy-evaluate-retrain workflow and says its Lab service embeds researchers and engineers with customers to help design training strategies, evaluation systems, data pipelines, and production deployments.

Fireworks reports that DeepSeek V4 Pro 0813 performed strongly in CyberGym, a Berkeley-developed benchmark based on real vulnerabilities in open-source projects that requires agents to identify flaws, create working exploits, and patch affected code. Across tested tasks, the company says DeepSeek produced tool calls without refusals or output truncation, entered validation on 97.3% of a 697-task common cohort, and achieved a 53.7% reward rate at an estimated $2.50 per solved task, outperforming GPT-5.5 on both reward rate and cost but trailing Kimi K3’s 68.4% solve rate. The evaluation emphasizes that long, multi-step cybersecurity tasks can expose limitations of models that refuse or fail to generate usable artifacts, with Claude Opus 4.8 showing a low completion rate in the reported comparison. Fireworks argues that DeepSeek’s main weakness is generating initial proof-of-concept exploits rather than applying patches after validation, suggesting that improved prompting or scaffolding could raise its results, and positions open models as a practical option for high-volume vulnerability auditing and remediation workflows.

Fireworks announced availability of DeepSeek V4 Pro 0813 through serverless APIs, dedicated deployments, and training tools, positioning it for long-horizon agent workloads with a 1 million-token context window, native tool calling, and retained reasoning history. In its evaluations against Kimi K3 and Fable 5, DeepSeek led on SWE-Bench Verified at 95.2% and LiveCodeBench at 92.0%, while Fable 5 performed better on Aider Polyglot and matched DeepSeek on Terminal-Bench. Fireworks reports that DeepSeek completed standard coding tasks at substantially lower cost per solved task than Fable 5, though it identified Java as a relative weakness that makes multi-model routing preferable for some codebases. An oracle-routing analysis combining DeepSeek and Fable 5 achieved higher aggregate accuracy and lower estimated cost than Fable 5 alone, although the company notes that oracle routing is a theoretical upper bound rather than a deployable policy. Fireworks also offers fine-tuning through SFT, DPO, and RFT, recommends using DeepSeek’s native harness or a minimal tool setup to avoid performance losses from interface mismatches, and cites internal security tests showing no refusals or output-length truncations across 840 adversarial runs.

Harvey and Fireworks announced Tenet, a model post-trained from Kimi K3 using asynchronous reinforcement learning to improve long-horizon legal-agent work. On the Legal Agent Benchmark, which requires completed legal deliverables to satisfy every grading criterion, Tenet achieved a 19.7% all-pass rate versus 10.8% for the base model and set a leading result on LAB Contracts, while also improving on unseen corporate-law and contract-redlining agent benchmarks. The model retained comparable performance on broader legal knowledge and reasoning evaluations and operated at nearly the same per-task cost as Kimi K3 despite completing almost twice as many LAB tasks. Fireworks attributes the training effort’s scalability and reliability to shared training-and-inference numerics, which reduce instability during long reinforcement-learning runs, and deterministic batch-invariant serving, which makes evaluations reproducible; the partners plan to publish a more detailed account of the reward design and training methodology.

Researchers applied the Jacobian Lens, a trained probe designed to map transformer hidden states to likely vocabulary, to the open-weight Kimi K3 and Qwen3.5-9B models to examine whether task-relevant concepts can be detected before they appear in generated text. Across paired-copy experiments and ten prompt episodes involving arithmetic, citrus, factual questions, rhyming, sports, code errors, and language tasks, the probe identified relevant token families in internal states even when visible outputs were identical or had not yet named the concepts. A separately fitted lens on Qwen reproduced 19 of 20 vocabulary-family matches originally defined for Kimi, while transcript-region tests found signals during prompt processing, chat-template context, and early response generation. The study notes that results are limited by tokenization, particularly for multi-token numbers such as “21” and “42,” and that fitting the probes required substantial compute. The authors conclude that open models can expose measurable internal representations associated with later outputs, enabling independent investigation of model behavior, though the findings rely on fitted vocabulary probes rather than direct proof of human-like reasoning.

Meta’s Muse Glimmer, now available through Fireworks, is a 30-billion-parameter dense multimodal model designed for always-on agents that must coordinate long, multi-step tool use, recover from failures, and maintain context across extended workflows. Its architecture combines 52 transformer layers, a roughly 1.8-billion-parameter image encoder, a 128K-plus-token context window, sliding-window attention with periodic global attention, and a small KV cache intended to reduce serving costs at high concurrency; it also supports DFlash speculative decoding for lower latency and is released under Apache 2.0. Fireworks positions the model for customer support, autonomous research and monitoring, and repository-scale coding, while recommending human approval for irreversible actions and tool-based retrieval for information after its January 2026 knowledge cutoff. Meta-reported results show Muse Glimmer outperforming Gemma 4 31B and Qwen 3.6 27B on several agent-oriented benchmarks, including MCP Atlas, DeepSearch QA, Gaia2, WildClawBench, and SWE-Bench Pro. Fireworks offers serverless and on-demand deployment, autoscales for variable agent traffic, and recommends settings including temperature 1.0, top_p 0.95, top_k 64, and high or xhigh reasoning strength for agentic and coding tasks.

Fireworks announced a partnership with Voyage AI by MongoDB that makes Voyage’s embedding, multimodal retrieval, and reranking models available natively on its inference platform alongside open models for post-training and serving. The integration is intended to let teams run embedding, retrieval, reranking, and generation through one API and infrastructure environment, reducing separate vendor connections, latency hops, and data-handling boundaries. The post argues that retrieval quality is central to AI systems built on proprietary data, and presents specialized post-trained open models plus data grounding through retrieval as complementary components of business-specific AI. Available Voyage models include voyage-4-large, voyage-4, voyage-4-lite, voyage-4-nano, voyage-multimodal-3.5, and rerank-2.5, with suggested applications ranging from customer support and internal knowledge assistants to agentic workflows, multimodal RAG, semantic search, recommendations, and deduplication. Fireworks states that Voyage 4 Large leads several competing embedding models in its cited average retrieval-quality comparison and directs users to its cookbook and account signup process to build an integrated pipeline.