Grok 4, xAI’s latest flagship LLM, is now available on Maxim. Access powerful capabilities like PhD‑level reasoning, a 256k token context window, and advanced math performance to supercharge your experimentation and evaluation workflows.
✅ Start using these models via the xAI provider on Maxim: Go to Settings > Models > Select xAI provider > Add Grok 4 to start using.
We’re excited to announce the removal of the 1MB size limit for log uploads! Previously, logs larger than 1MB were truncated or rejected, but now you can push logs of any size directly into Maxim — perfect for handling large traces, detailed debug output, or verbose agent sessions.
What’s changed?
This unlocks deeper observability for complex workflows and ensures you never lose critical context due to log size constraints.
You can now test, automate, and scale your workflows more easily than ever. Simulate multi-turn interactions with your prompt and assess its performance in scenarios with user follow-ups.
Now you can bring your own prompts, connect MCP tools, integrate RAG pipelines, and launch thousands of scenario simulations in Maxim AI with a single click.
We've added support for scheduling automated runs on HTTP endpoints on which your AI Agent is deployed. You can now configure periodic simulations by providing an HTTP endpoint, attaching a dataset, and choosing one or more evaluators. The system will trigger these simulations at the defined interval, enabling continuous performance tracking and regression detection over time, without manual intervention.
Improved Jinja2 Parsing: We've enhanced our Jinja2 parser to deliver more precise variable extraction and cleaner, more reliable template rendering.
Flexible Column Type Editing: You can now modify the data types of columns in your datasets (excluding file-type columns), giving you greater control over schema customization.
We're excited to announce the public release of Bifrost, the fastest, most scalable LLM gateway out there. We've engineered Bifrost specifically for high-throughput, production-grade AI systems, and we've optimized performance at every level:
🔸 ~0 heap allocation during live requests (configurable)
🔸 Actor pattern to avoid fetching the config at request time
🔸 Full use of Go’s concurrency primitives
🔸 Lightweight plugin system to keep the core minimal
🔸 Support for multiple transport protocols (HTTP, gRPC)
And it’s:
🔹 Open source
🔹 Written in pure Go (A+ code quality report)
🔹 40x lower overhead (based on LiteLLM’s published benchmarks)
🔹 9.5x faster, ~54x lower P99 latency, and uses 68% less memory than LiteLLM
🔹 Built-in Prometheus observability
🔹 Plugin store for easy extensibility
Check out our GitHub repo to get started. Read more about Bifrost benchmarks in our blog.
We're excited to announce our single line integration with Mistral – enabling fast, efficient, and observable LLM workflows, powered by Maxim's eval and observability stack. Here's what you get out of the box:
Here's what you get out of the box:
Traceability for LLM calls: Capture complete request-response traces with prompts, parameters, and metadata across all your Mistral calls.
Run evaluations: Run custom or out-of-the-box evals on every Mistral interaction, including multi-turn chains, to measure quality, coherence, safety, and more.
Usage & performance dashboards: Track latency, cost, token usage, and output metrics in real time.
Smart alerting: Set up Slack/PagerDuty alerts based on spikes in latency, failure rates, cost thresholds, or feedback scores.
📘 Get started now: Mistral + Maxim Integration Docs
We've enhanced our data export functionality with built-in protection against CSV injection (formula injection) vulnerabilities. You can now safely export logs, datasets, and other data in CSV format, ensuring that even untrusted input won't trigger formulas or pose security risks when opened in spreadsheet tools like Excel or Google Sheets.
The Expected Tool Calls column allows you to specify the tools you expect an agent to use in a scenario, ensuring the AI agent is choosing and invoking the correct tools as part of its reasoning process.
This column has been improved with the addition of powerful combinators for flexible evaluation of agent behavior:
inAnyOrder: Allows you to specify multiple tool calls that can be executed in any sequence while still being considered valid. This is perfect for scenarios where the order of operations doesn't matter.
anyOne: Enables you to define alternative tool calls where any single one satisfies the requirement. This is ideal for cases where multiple approaches can achieve the same outcome.
This provides greater flexibility when evaluating agent behavior, particularly in complex scenarios with multiple valid solution paths. Learn more about Expected Tool Calls column .
You can now define a "Conversation History" column in your test datasets to include prior multi-turn interactions between the user and LLM alongside your "Input" while running prompt tests.
Passing chat history to the LLM provides critical context, enabling it to understand the ongoing dialogue rather than treating each input as an isolated query. This allows you to simulate real-world interactions more accurately. Here’s how it works:
Read more about the Conversation History column.
