AI news, read from an operations desk

Latest

Prismus raises €163K from Venture Kick for Lupa AI platform

Prismus, a deeptech startup, received €163,000 (CHF 150,000) from Venture Kick to develop Lupa, an AI platform that reads research, patents and market data and ranks likely markets, customers and partners. It reports thirty paying customers. The money goes toward deep tech markets and individual users.

What changes for operators — If your sales or partnerships team spends hours each week hunting for target accounts, collaborators or niche markets across scattered sources, Lupa is an example of a tool built to do that ranking automatically with sources you can check. The source does not give pricing, integrations or a launch date for new users, so treat it as a product to watch rather than one to buy today. The practical test for any such tool is whether each recommendation links to a source your team can verify before anyone acts on it. For a 10-200 person company, that decides whether the output goes straight into a prospect list or becomes one more thing to double-check by hand.

Google plans one robots.txt token for AI training opt-outs

If you run a B2B company with a public website, the content, docs and knowledge base you publish are what AI developers may train on. Today that choice is made provider by provider. Have whoever owns your site or marketing ops list what your robots.txt says now, and decide whether you want a blanket opt-out or a mix. Keep in mind that allowing Google-Extended specifically makes a site eligible to appear in Gemini Apps responses and receive clicks via grounding, so a blanket opt-out may cost you visibility there. Nothing needs to change today, since Google gives 2027 as its target, but the decision is cheap to make now and easy to forget later.

Serval, Hex and General Compute raise AI rounds

If you run sales, support or operations at a 10-200 person company, treat these rounds as a vendor-risk input, not a buying signal. Serval's coding agents, Hex's analytics platform and General Compute's inference cloud all now have money to keep building. Before you wire any of them into a workflow, ask for a pilot on one real process, such as ticket triage or a weekly pipeline report. Get the pricing in writing and confirm you can export your data and logic if the vendor changes direction.

Microsoft puts a managed agent harness into Foundry preview

If you run sales, support or operations at a 10-200 person company and have a pilot agent that stalls on multi-step work, this lowers the engineering cost of keeping it running, because you stop maintaining the loop and sandbox yourself. The catch is that skills, approvals and tool permissions are still yours to define. Use the free period to test one workflow, such as a weekly account research brief, and log how many sandbox hours it uses. Any agent that writes to a CRM or helpdesk needs a check for existing records so a restart does not create duplicates, and commercial decisions should stay with a person.

Mistral Large 4 ships as open weights on Hugging Face

If you run sales, support or operations at a 10-200 person company, this model is mostly a hosted option for you, not a box to rack. The reference deployment on Hugging Face's Inference Endpoints uses eight NVIDIA H200 GPUs, so renting by the hour or using serverless Inference Providers is the realistic route. Check two things before a pilot: the license restricts selling the model as a service, which matters if you resell automation to clients, and reasoning cannot be turned off, so expect every reply to carry reasoning overhead in latency and cost. Test it on your own tickets, CRM notes and documents before trusting the vendor-reported benchmarks.

AI broker startup i23 raises £4.9m seed round

There is nothing to buy or switch on yet: the source reports funding and a stated direction, not a released product or pricing. For anyone running operations at a 10-200 person B2B company, the useful signal is where investors expect AI to take over first: routine administrative work in a fragmented, intermediary-heavy market. If your own team spends its week on repetitive admin around clients and paperwork, list those steps now and measure their volume, so you can judge any vendor in this space against real numbers. If you buy insurance through a broker, expect brokers to start talking about AI-handled admin, and ask what stays with a human.

OpenAI adds API credits to its $100-$500 ChatGPT plans

If someone on your team already holds a $100, $200 or $500 ChatGPT plan, check whether that subscription now carries API credits and which API organization they are applied to. Use them for internal prototypes, such as a lead-qualification script or a support-ticket triage test, before anyone approves a separate API budget. Treat them as a trial allowance and confirm the expiry terms on OpenAI's help page before building anything you depend on, because the source does not state them for OpenAI.

Two Swift packages bring open speech and decision models

This matters only if your team builds or commissions Mac or iOS software. A support tool that transcribes a call, tells the speakers apart and routes the ticket to billing, technical or sales can now run on the agent's own Apple silicon machine, with the model returning a probability and a confidence value for each routing choice. That lets you set a review threshold, for example sending low-confidence tickets to a person. The source warns that confidence is not correctness, so measure the error rate on your own tickets and audio first. If your stack is web, Windows or Android, nothing here changes your setup yet.

AWS Batch adds quota shares for shared GPU training on SageMaker

This matters only if your company trains or fine-tunes models on GPUs that several people or projects share. If a 10-200 person team fits that description, the quota shares replace the Slack threads about who holds the GPUs: give each project a share, mark which ones may lend or borrow, and let the scheduler reclaim capacity. If you only call hosted models through an API, there is nothing to change. The source gives no pricing or regional availability, so check the AWS Batch user guide before you plan around it.

Machenta emerges from stealth with $40M seed for robot learning

If you run operations at a 10-200 person B2B company, there is nothing to buy or integrate from this announcement: Machenta has named no product, price or customer. It matters only if your work involves physical steps such as packing, assembly or tidying, where robot setup cost and time have been the barrier. Note the claim and check back for a pilot or pricing; until then, keep automation budgets on the software workflows that exist today.

AWS adds team access and health metrics to Nova Act QA Studio

If you run a web product and have been piloting AI-written tests, this update addresses what breaks once a second and third team join: nobody knows whose tests are failing, and everyone can edit everything. A 10-200 person company would map each product to one application and one owner, give developers Editor, give release managers Executor and leave management Viewer. The flaky-test ranking shows which tests to fix first. Note that it is a reference solution you deploy and maintain with AWS CDK in your own account, so it only suits you if you already run on AWS and have someone to own the deployment.

Salesforce to buy Rune Technologies for defense logistics AI

If you run sales, support or operations at a 10-200 person B2B company, nothing in your Salesforce setup changes because of this deal: it targets military logistics and has not closed. What it shows is where the vendor is putting AI agent effort, namely systems that keep working offline and sync back to a central record. If you rely on field teams, warehouses or sites with unreliable connectivity, ask your vendors how their agents behave when the network drops and how conflicts get resolved afterwards. Otherwise, treat this as vendor-strategy context and check your roadmap only if you depend on Salesforce's government products.

IBM updates Granite Switch: new ASR model, Decision adapter

If you run sales calls or a support line, the speech model matters most: transcripts arrive with punctuation and capitalization, so you can pass them straight into summarization or routing steps without a cleanup pass. Because 4.1.1 is backwards compatible and leaves the base model alone, a team already on 4.1 can swap in the new adapters and re-run its own call samples, with no migration project. The source gives no pricing, licence or accuracy figures for the new speech model, so test it on your own recordings and languages before relying on it. The Decision adapter gain is reported only against IBM's own test set, so treat it as unverified for your workflows until you have checked it.

Mistral publishes Large 4 weights under a modified MIT license

If you run sales, support or operations at a company of 10-200 people, the practical change is a new option for agents that write code, read documents or browse the web. You can test it through the API at $0.70 input and $2.10 output per million tokens, or through Microsoft Foundry, before deciding anything about hosting. Self-hosting a 1T-parameter model is a separate infrastructure decision, and the source gives no hardware requirements, so ask for them before you budget. The license matters only if you resell model access as a service. Run your own tasks, such as ticket triage, CRM data cleanup or report generation, through the low, medium and high reasoning modes and compare cost per finished task, because the vendor's benchmarks do not cover your workflows.

Vera Health raises $28.5M Series A for clinical AI search

If you run support or operations at a 10-200 person B2B company, Vera is not a tool you will buy, but its design is a useful checklist for any AI answer system you deploy. Every answer links to its sources and carries an evidence-strength label, and conflicting sources are shown openly. Ask the same of your own support bot or internal search: can an agent see where an answer came from and how reliable it is? Vera's revenue model is also worth noting. Individual users adopt it free, and the institution pays once usage is widespread. A pilot with a few power users in your team can build the case for a paid rollout.

Google takes ads in AI Mode international for the first time

If your team buys Google ads to feed the pipeline, a new placement is about to appear in front of longer, more specific queries. The source gives no rollout dates, country list or buying controls, so do not move budget yet. Do three things now: check whether your ad reports can separate AI Mode traffic from classic search results, make sure landing pages answer detailed questions directly, and track lead quality by source in your CRM. Sales and support teams will then have evidence on whether the new traffic converts, and the first numbers will come from your own account rather than from Google's claims.

Itaú buys AI model startup NeoSpace for nearly R$3 billion

For a 10-200 person B2B company, the useful part is the bottleneck Itaú describes, not the price. Its in-app chat engine is still limited because the knowledge sits in employees' heads and has to be gathered through interviews and turned into digital form. If you plan a support or sales assistant, start by writing down how your team answers questions, which products and terms matter and what the customer record contains. A model, bought or rented, is only as useful as that material.

Anthropic's S-1 says it may compete with app startups

If a sales, support or operations workflow at your company runs on a niche AI app built on a third-party model, ask the vendor two things: which model provider it depends on, and what its plan is if that provider ships a competing product. The filing names no categories, so nobody can rule any out. Keep your data exportable, avoid long prepaid contracts with young vendors where you can, and keep the workflow logic in a layer you control so the model or the app can be swapped.

DeepMind opens Genie 4 world model to trusted testers

Nothing here changes a sales, support or operations workflow today: access is limited to trusted testers and a waitlist, and the Gemini API route is only for early access partners. The part worth tracking is the testing idea. DeepMind pairs Genie 4 with Gemini models so an AI system can be tried in a simulated environment before it touches the real one. If you plan to put agents into live processes, watch whether that kind of pre-deployment testing reaches the API. Until then, treat it as a signal and not as a purchase.

Liquid AI releases open-weight LFM3 family of 2B-class models

If you run support or back-office operations, the useful parts are the d1 decision models and the fine-tuning notebooks. d1 answers several questions about one ticket or invoice in a single pass, each with a probability, so routing, refund detection and urgency scoring need no separate prompts. The notebooks cover turning messy documents and receipt photos into JSON. Because the models are small and open-weight, a short fine-tune on your own tickets and terminology can run on private infrastructure. Treat every benchmark as Liquid AI's own and test on your data before relying on it.

Strawberry takes its AI browser out of beta for business teams

If you run sales or support at a 10-200 person company, Strawberry is a cheap way to test whether a non-developer on your team can build and share a working agent, such as a prospecting assistant, without engineering time. Start on the free tier with one narrow task and watch how fast credits are used before choosing among the $20, $85 and $250 plans. Because the agents run on the user's device and the product is model agnostic, check what that means for your data handling and for which models you end up paying for. The source gives no customer results beyond early prospecting assistants, so treat it as a pilot, not a replacement for your current process.

Hugging Face libraries now run natively on TPUs via TorchTPU

If your sales, support or ops automation calls a hosted model API, nothing changes for you today. It matters if you self-host open models for a support assistant, document processing or classification and are weighing hardware. TPU on Google Cloud can now be benchmarked with the same Hugging Face code you already run, by passing "tpu" as the device. That makes a trial a small engineering task rather than a porting project. Treat the speedups as the vendor's own numbers, and run your own model, batch size and latency target before moving anything.

Reduck raises $3M pre-seed to speed up browser-based AI agents

If your team still has people downloading invoices from vendor portals, copying data out of a customer portal, or keying records into a CRM that has no API, this is the category to watch. Scripted browser access is the approach Reduck is betting on to let an agent handle that work. Today the evidence is a vendor's own claim and a pre-seed round, so treat it as something to test on one narrow, repetitive task, such as pulling invoices from a handful of suppliers, rather than something to rebuild a process around. Measure time per task, cost per run and failure rate against your current manual routine before committing.

Indurex lands €7M for AI-driven industrial cybersecurity

If you run operations at a 10-200 person company outside industry, this changes nothing for you today. If you run a plant, a manufacturing line or energy assets, it is a signal that security tooling is moving from watching network traffic to watching engineering and process data, and that mid-sized operators are named among Indurex's early customers. The source gives no pricing, packaging or availability dates, so the practical step is to ask your current OT or IT security vendor whether it sees engineering configurations and safety systems, not only the network.

NVIDIA adds StormScope EU nowcasting model to Earth-2

If you run operations at a 10-200 person company in energy services, logistics or insurance-adjacent work, you will probably not train this model yourself. The realistic route is through a weather vendor: Vaisala Xweather says it is adapting StormScope with lightning data and exposing the warnings through its SaaS products and AI-assistant integrations. So the useful question is for your current weather-data supplier: does it use regional nowcasts, and can the alerts feed your dispatch, scheduling or customer-notification workflows? If local conditions do not affect your delivery or service commitments, this item needs no action.

Canva launches Craft, its own image model, for all users

If your sales or support team already builds decks, social posts or one-pagers in Canva, Craft puts generation inside that workflow, and outputs can be turned into layered designs with Canva Magic, so staff edit rather than regenerate. The part worth watching for automation is the announced Canva API access and the planned MCP support for ChatGPT and Gemini: that would let a workflow request on-brand visuals without a person opening the editor. None of it is priced in the source, so test quality on your own templates before routing any process through it.

Amazon Polly upgrades Generative voices, adds Hindi and Finnish

If your support line, order-status IVR or voice agent already uses a Polly generative voice, the new engine arrives on its own and costs nothing extra, so the work is checking that your own scripts still sound right. Listen to the calls where the voice reads back order numbers, addresses, totals and confirmation codes, since that is where AWS says accuracy improved. If you serve Finnish or Hindi speakers, or Korean, German or Brazilian Portuguese ones, the catalog now has generative options to test against what you use today. If your flow depends on speech marks, those are still not available on generative voices, so check that before you switch. Before deploying across Regions, confirm the voice you picked exists in the Region you run in.

Microsoft releases ThinkingBox-Training for agent RL tuning

You will not run a 24-GPU training job at a 10-200 person company, but the post gives you a better buying test for agents that write to your CRM, ticketing or billing records. Ask for the share of tasks passed on every repeated attempt, not just average accuracy, and grade by the final state of the record rather than by what the agent says it did. If you have a vendor or partner tuning agents for you, ask whether the reward comes from a check on the resulting data or from a model's opinion of the transcript.

AWS SageMaker adds training session APIs for custom LLM loops

For a 10-200 person B2B company, this matters only if someone on the team already fine-tunes models, for example to train a support-reply model against your own quality checks or a sales-qualification model against your own scoring rules. In that case you no longer have to choose between a fixed managed recipe and running a GPU cluster, and per-token billing makes experiments cheaper to compare. If you only call hosted models through an API, nothing changes for you. The first step is to run the GSM8K sample with your own reward function and see what a run costs before committing to anything.

IBM Research ships a four-model Granite forecasting ensemble

If your team forecasts anything as a time series, such as ticket volume, order flow or staffing load, the practical change is one less model-selection decision. You load a single recipe, get a median plus a 0.1 to 0.9 range, and use the range for planning. Before trusting it, backtest on your own history and compare it with the best single member. Check that the 80% interval actually covers about 80% of outcomes, and measure runtime, because four checkpoints load instead of one. For a small team without a data scientist, this is a reasonable first forecaster to test, not a proven winner.

US banks restrict third-party AI agents from customer accounts

If anyone on your finance, sales or operations team has given an AI agent a bank or brokerage login to pull balances, reconcile or move money, plan for that to stop working. The banks' terms require a human to log in, and they cannot yet tell a legitimate agent session from an unauthorized one. Audit where credentials have been shared with agents, move those workflows to read-only access of the kind expense-tracking apps already use, and keep a named person approving every outgoing payment. A bank may also hold you responsible if something goes wrong after you shared your credentials with a bot.

Walmart opens an MCP gateway for AI shopping agents

If you run sales or ops at a 10-200 person B2B company, Walmart's setup is a checklist for any system you let an agent touch: give each agent its own identity so its actions are visible apart from a person's, cap what it can spend or commit, decide which actions need a human approval, and write down terms that justify cutting off an agent that oversteps. It also shows the choice you will face if customers or partners start sending agents to your order forms, quote tools or support portals: block them outright, as Amazon has, or give them a defined, controlled entry point. Making that decision deliberately beats discovering agent traffic by accident.

OpenAI reportedly seeks $30B at $1.25T valuation

Nothing in these reports changes what an OpenAI-based workflow costs or does today: no product, plan or price is mentioned. For a 10-200 person B2B company the useful move is housekeeping. List which sales, support or ops automations call OpenAI directly, and check that the model call sits behind one layer you control, so a provider can be swapped without rebuilding the process. Treat the valuation figures as unconfirmed reporting about the vendor, not as a reason to change a contract or a roadmap.

Microsoft brings Foundry Local to disconnected Azure Local

If you run operations at a 10-200 person B2B company, you will probably not deploy Azure Local yourself, and nothing here changes your tooling today. It matters if your buyers are in government, critical infrastructure or regulated industries and ask where models run and who can touch the data. Microsoft is now describing a packaged route for local AI over internal documents, which could shape procurement questions and security questionnaires. Keep a short answer ready on where your automations run and what data they see.

Commerce.com turns on WebMCP for nearly 100,000 storefronts

If you sell through a Commerce.com store, agent access is now on by default, so check the opt-out switch in your dashboard and decide deliberately. If you run your own site on Cloudflare, the same edge WebMCP feature has been in beta since August, so you can test it without touching your theme. Treat it as a pilot: WebMCP sits behind experimental Chrome flags and in a few agentic browsers, so little real traffic will use it yet. The useful question for sales and ops is whether your catalogue, pricing and cart logic are clean enough for an agent to call, because the checkout step stays with the human buyer.

Vivenda leads €1.5M round in food-waste AI maker Ganiga

For a food-service or catering company of 10-200 people, the practical point is that waste can now be measured by cameras above the bins, without weighing, and fed into purchasing and menu planning. Ganiga says its system already runs in operations serving more than 30,000 meals a day. If you run such a kitchen, the question is whether your ordering decisions rest on measured waste or on guesses. Ganiga also plans to automate ingredient ordering, but that is a stated plan, not a shipped feature, and the source gives no pricing.

Anthropic's Claude 5.5 family: Opus, Sonnet and Haiku released

If your support or sales workflows already run on Claude Opus 5 or Sonnet 5, these releases are a prompt to re-test the same tasks on the 5.5 versions and compare cost per ticket or per lead summary. Anthropic describes Haiku 5.5 as built for summarization, subagents and browser use, so it is the candidate for bulk steps like call summaries or inbox triage. Use your own tasks as the benchmark, because the cost figures are Anthropic's own and depend on the workload.

IBM's ALTK-Evolve tops τ²-bench retail with GPT-4.1 memory

If you run support on a 10-200 person team, the useful point is where the gain comes from: a maintained rulebook in the system prompt, not a bigger model. Rules such as authenticating before any order action, confirming changes explicitly and stating refund timing are the kind your team already writes down. The price is input tokens, about 65% more per retail conversation with the full 292-rule set, and a smaller rule set can be the more reliable choice in some domains. Test on your own transcripts, since the evaluation here is small and run on simulated customers.

OpenAI absorbs Orchestra Bio into its Rosalind life-science work

If part of your sales, support or operations stack runs on a niche AI vendor, this deal is a prompt to check your exposure. Orchestra's agents did back-office work such as tracking finances and sorting out scheduling, and that product line is now being folded into a larger lab's offering. For a 10-200 person company, the practical steps are to read the change-of-control and data-handling terms in your vendor contracts, confirm you can export your workflows and data, and keep a fallback for any process that depends on a single small AI supplier. Nothing in the source says Orchestra's existing customers are affected, so treat this as a due-diligence reminder, not a warning.

Slack adds Frontline, a mobile workspace for non-desk staff

If you run operations at a company of 10-200 people with shift workers, technicians or store staff, the useful part is that handoffs, alerts and escalations can move out of group texts and into governed channels that agents and workflows can act on. The catch is the entry requirement: Frontline is an add-on to Slack Enterprise+, and the source gives no price. Check your current Slack plan and ask sales for a quote before planning around it. Nothing in the announcement shows how it performs outside the vendor's own scenarios, so a pilot with one site or one shift is a sensible first step.

Cloudflare previews Clairic, an extraction model on Workers AI

If your support or finance team reads order numbers, dates, amounts and damage descriptions out of inbound emails, receipts or invoices, this is a cheaper way to fill those fields than prompting a general LLM for JSON. The confidence score and null probability give you a rule for routing: auto-process high-confidence values, send low-confidence ones to a person. The estimate and forecast fields could also feed lead scoring or volume planning. These are Cloudflare's own examples, so run it on a sample of your real documents before wiring it into anything that moves money.

OpenCode Console breach exposes API keys of 1.4% of users

If anyone on your team pasted provider API keys into a hosted workspace of a coding or automation tool, treat those keys as a liability you do not control. Check whether your team used OpenCode Console, and if you did not get an email from OpenCode, the company says you were not affected. Either way, list every place a production key lives, set spend alerts on each provider account, and give each key one job so it can be revoked without stopping everything else. The users who caught this fastest were the ones whose spend alerts fired overnight. A 20-person company with one shared key in a workspace has no way to separate its own usage from an attacker's until the bill arrives.

Amazon Bedrock adds TwelveLabs video model Pegasus 1.5

For a 10-200 person B2B company, this matters only if you hold long video: recorded demos, support sessions, training material or webinars. The source's own examples come from media and sports, so any fit with sales or support recordings is our reading, not the vendor's claim. If you do have such a library, you can define a segment once, such as a pricing discussion or a product shown on screen, and get consistent JSON for every file. That feeds a CRM, a ticket queue or a search index without anyone scrubbing through footage. If you have no long video, there is nothing here for you to act on. Check the Region and the separate AWS Marketplace billing before you plan around it.

Together AI launches Post-Training for custom open-model tuning

For a 10-200 person B2B company, this is relevant only if one narrow, high-volume task, such as ranking leads or scoring candidates, is costing too much on a closed model. Jack & Jill's case in the source is that shape: an open model matched a closed model in offline tests and is cheaper to serve. Billing is per GPU-minute, so budget an experiment before you start. Without ML staff, you would lean on Together's forward-deployed engineers, which the source offers as a partner option.

Meta releases Muse Spark Open, an 816B open-weight model

For a 10-200 person B2B company, the useful fact is that a frontier-scale model with tool calling and image, audio and video input is now open-weight, so support or sales workflows that handle screenshots, call recordings or screen captures could run without a closed API. The cost is hardware: the smallest validated hosted setup in the source is eight GPUs, so most teams will not self-host. The realistic route is a managed endpoint or a pilot. Test one narrow workflow, such as ticket triage with screenshots or call transcription, and compare quality and cost per task against what you use today. Because reasoning cannot be turned off, budget for the extra tokens.

Ambio raises €5.1M seed for AI recycled-plastics buying

Ambio's sales pitch is a pattern any ops lead at a 10-200 person company can check against their own purchasing or supplier work. Its supplier information lived in emails, phone calls and WhatsApp, so the company first built a structured dataset, then automated the matching and document reading, and kept people on the commercial decisions. If your team buys from fragmented suppliers, that order is the practical takeaway: fix how specifications, pricing and outcomes are recorded before expecting an agent to do much. The product itself only matters if you buy recycled plastics; the round is a signal of where agentic procurement is being funded, not a tool most readers can wire in today.

Replit switches Agent default to OpenAI, Claude still available

If anyone at your company builds internal tools, prototypes or automations in Replit, check which model each project actually uses, because the default is what most people end up on. Pick a model deliberately for anything customers or colleagues rely on, and rerun your key workflows after a default change to catch shifts in output quality. Also treat this as a reminder not to tie a process to a single model vendor: Replit calls itself model agnostic, and your own stack should be too. Watch the vendor's changelog for confirmation, since the Anthropic side is still disputed.

Replit switches its default model from Anthropic to OpenAI

If your team builds internal tools, prototypes or automations on a platform like Replit, the model underneath is the vendor's choice unless you set it yourself. Most users never change the default, so a vendor-side switch can alter output quality, cost and behavior without any action on your side. Check which model each workflow actually runs on, select one explicitly where results matter, and keep a small set of test cases you can rerun whenever a default changes.

Intuit buys AI billing startup Sequence, reportedly for $200M

If your invoicing or revenue process runs on Sequence, the source says nothing about changes to the product, pricing or contracts, so ask your account contact in writing what the acquisition means for your plan and support terms. If you use any small AI billing or payments tool, record who owns it, when your contract renews and how you would export your data. An ownership change is the moment to check that, before it becomes urgent.

Reflection releases Beam, a 501B open model under Apache 2.0

For a 10-200 person B2B company, Beam is a model to test through an API, not to host yourself. Self-hosting means roughly 280 GB of weights even in the 4-bit NVFP4 format, and the post suggests four B200 GPUs. Hosted access through Baseten, Fireworks or Together lets an ops lead run it on real tasks first, such as ticket triage or order checks. Treat the numbers as vendor-supplied. The tool-calling scores are modest: Tau3 Banking is 36.1% and AutomationBench (public split) is 39.3 for the BF16 checkpoint. Run your own workflow before replacing a current model.