# Meridian Air CivoE **Published:** November 4, 2026 **Source:** [webAI](https://www.webai.com/research/meridian-1-civoe) --- ## Same foundation. A different level of capability. Specialization substantially changed what the model could accomplish on the tasks we evaluated. ### Experts lift the same foundation from 59.8 to 85.5 Composite score on the same evaluation items: the Meridian Air foundation on its own scored 59.8 out of 100; Meridian Air CivoE scored 85.5, an improvement of 25.7 points. Up 25.7 points without changing the foundation Measured on a suite curated around specialized tasks. The improvement does not mean every task improves by that amount, or that it would hold on unrelated workloads. Figure 2 · Composite score out of 100, same evaluation items Equal-weight average of task-family scores, each normalized to a 0–100 scale. Metrics include accuracy, F1, TEDS, and other task-specific measures. The overall figure is a composite score, not a universal accuracy percentage. 14 source benchmarks · 12B foundation model. | Configuration | Composite score (0–100) | | --- | --- | | Meridian Air | 59.8 | | Meridian Air CivoE | 85.5 | Both configurations use the Meridian Air foundation. CivoE adds its experts and selects the right one for each task. The 25.7-point improvement was measured across the same selected evaluation items. It does not mean every task improves by that amount, or that the increase would necessarily hold across unrelated workloads. The result demonstrates the potential of specialization within the tested domains. ### How does a specialized 12B system compare with frontier AI? The most useful way to answer that question is to evaluate the models on the same problems. We tested Meridian Air CivoE against frontier and open-weight models using selected items drawn from 14 public benchmark datasets. The score covers three kinds of everyday work AI agents rely on: - Reads documents Contracts, financial reports and passages - Understands images Photos, pictures of tables and detailed scenes - Sorts and protects information Personal data, customer requests and security flaws The results show that a compact, specialized system can achieve frontier-competitive performance on certain categories of work. ### Level with Gemini 4, ahead of Claude Opus 5.5 and GPT-5.6 Composite scores out of 100 on the same evaluation items: Meridian Air CivoE 85.5, Gemini 4 84.7, Claude Opus 5.5 82.7, GPT-5.6 71.0, Meridian Air 59.8. A 12B system with experts scores alongside frontier models Frontier models were accessed through Cursor agent workers. Exact model revisions and reasoning settings were not fully controlled. These scores are not official leaderboard submissions or definitive rankings of general intelligence. Figure 3 · Composite score out of 100, same evaluation items Composite score: equal-weight average of task-family scores, each normalized to 0–100. | Model | Access | Composite score (0–100) | | --- | --- | --- | | Meridian Air CivoE | Same foundation, with experts | 85.5 | | Gemini 4 | Via Cursor agent | 84.7 | | Claude Opus 5.5 | Via Cursor agent | 82.7 | | GPT-5.6 | Via Cursor agent | 71.0 | | Meridian Air | Foundation only, without experts | 59.8 | Small gaps can be noise. To check which differences hold up, we resampled the evaluation items and measured how far each gap could move. ### Only the gap to Gemini 4 is too close to call Difference between Meridian Air CivoE and each comparator, with 95% bootstrap intervals: Gemini 4 +0.8 (−1.5 to +2.9); Claude Opus 5.5 +2.8 (+0.5 to +5.0); GPT-5.6 +14.5 (+12.1 to +17.1). The Gemini 4 interval crosses zero, so that difference is inconclusive. Ahead of Claude Opus 5.5 and GPT-5.6; tied with Gemini 4 These intervals do not correct for the earlier selection of task families where CivoE was already competitive. Figure 4 · CivoE's lead over each model, in composite points, with 95% intervals Bars show 95% bootstrap intervals; the dashed line marks no difference. An interval that crosses it is inconclusive. Frontier models were accessed through Cursor agent workers. Exact model revisions and reasoning settings were not fully controlled. These scores are not official leaderboard submissions or definitive rankings of general intelligence. | Comparison | Meridian difference | 95% bootstrap interval | | --- | --- | --- | | Gemini 4 | +0.8 | −1.5 to +2.9 | | Claude Opus 5.5 | +2.8 | +0.5 to +5.0 | | GPT-5.6 | +14.5 | +12.1 to +17.1 | ### Clear leads on personal data and photos; frontier models ahead on four families For each task family, Meridian Air CivoE's score minus the strongest tested frontier score: Personal-data tagging +15.0, Photo questions +14.0, Contract-clause finding +1.9, CVE-to-weakness mapping +1.0, Chess-board reading −1.0, Yes/no reading −4.0, Table-image transcription −5.5, Banking intent classification −7.0, Financial-report questions −11.0. Largest gains: personal-data tagging (+15.0) and photo questions (+14.0) Family-level differences have not been tested for statistical significance. Differences under 2 points are drawn in the lightest neutral and should not be read as a lead either way. Figure 5 · CivoE's score minus the best frontier score for each task family, in points Each row compares CivoE with the strongest frontier comparator for that family. Several tasks are custom evaluations built from public datasets with modified prompts, labels or scorers. | Task family | What it measures | Meridian Air CivoE | Best frontier comparator | Difference | | --- | --- | --- | --- | --- | | Personal-data tagging | Find and label personal information in text using a specific structured extraction format. The largest difference occurred on Nemotron PII; the advantage on Gretel PII was considerably smaller. | 91.8 | 76.8 (Gemini 4) | +15.0 | | Photo questions | Answer questions about photographs. Scores come from selected validation items and a filtered test subset under the lab's protocol, not an official VizWiz leaderboard ranking. | 74.9 | 60.9 (Gemini 4) | +14.0 | | Contract-clause finding | Locate the clauses in a contract that answer a specific question. | 90.4 | 88.5 (Claude Opus 5.5) | +1.9 | | CVE-to-weakness mapping | Map a published vulnerability description to the weakness category it belongs to. | 77.0 | 76.0 (Gemini 4) | +1.0 | | Chess-board reading | Read the position of every piece from an image of a chess board. | 98.0 | 99.0 (Gemini 4) | −1.0 | | Yes/no reading | Answer a yes-or-no question about a short passage. | 83.0 | 87.0 (Gemini 4) | −4.0 | | Table-image transcription | Transcribe an image of a table into structured markup. | 87.5 | 93.0 (Claude Opus 5.5) | −5.5 | | Banking intent classification | Classify a customer banking request into the correct intent. | 93.0 | 100.0 (Gemini 4) | −7.0 | | Financial-report questions | Answer questions that require reading and calculating from financial reports. | 74.0 | 85.0 (Claude Opus 5.5 and Gemini 4 (tied)) | −11.0 | ### Understanding the two largest task-family advantages #### Personal-data detection Our specialized system reached 91.8 on the combined personal-data evaluation, compared with 76.8 for the strongest tested frontier comparator. This matters because recognizing sensitive information is a critical step in many enterprise document workflows. The evaluation was built from Gretel PII and Nemotron PII datasets using a specific structured extraction task. The strongest difference occurred on Nemotron PII; the advantage on Gretel PII was considerably smaller. Further testing is needed across new document types, entity taxonomies, and operational settings. #### Photo understanding Meridian scored 74.9 across the selected VizWiz evaluation items, compared with 60.9 for the strongest tested frontier comparator. This suggests that targeted specialization can improve image question answering for a particular user need. The reported scores come from selected validation items and a filtered test subset under the lab's evaluation protocol. They are not an official VizWiz leaderboard ranking. Real-world reliability, accessibility outcomes, and performance across new photographs still need to be measured. ## What happens when intelligence scales through specialization? These results matter beyond one model or one benchmark. They suggest six directions worth investigating. ### The biggest model isn’t always the best model for the job. For much of the industry’s development, capability improvements have been associated with larger models, more training compute, and increasingly expensive infrastructure. Those methods remain important. But our results show that a 12B foundation, equipped with task-specific expertise, can achieve performance comparable to much larger general-purpose systems on selected workloads. This doesn’t mean small models have caught up on every dimension of intelligence. It means model size alone cannot tell us which system will be most useful for a particular task. Diagram comparing both systems after 1, 4 and 12 capabilities. They start the same size. The general-purpose model grows larger at each step. The Meridian Air foundation stays the same size while 1, 4 and then 12 small experts are added around it. General-purpose model · Grows with every capability Meridian Air CivoE · Same 12B foundation; experts are added 1 capability 4 capabilities 12 capabilities ### Intelligence may grow through expertise, not only parameters. Today, building a more capable AI system often means improving or replacing its foundation model. Specialization offers a different possibility. A shared foundation can support a growing civilization of experts, each developed to improve a particular capability. This suggests a path toward systems that acquire a broader collection of useful skills without requiring an entirely new foundation model for every additional task. Our research demonstrates substantial improvement across selected specialized tasks. It does not yet establish how this architecture performs with hundreds or thousands of experts. That is one of the central questions we intend to investigate. Diagram: the same field of points, shown once for each kind of work: documents, photos and contracts. Each kind of work pulls the points into a different arrangement; the number of points never changes. Shaped for documents Shaped for photos Shaped for contracts ### More useful work from the compute we already have. The cost of intelligence isn’t determined solely by how much intelligence a model contains. It also depends on how much computation is required to solve the task. A compact foundation with task-specific experts could make some repeatable workflows more economical than sending every request to a large general-purpose model. The potential is especially important for high-volume enterprise tasks, where small differences in cost per successful operation compound across millions of requests. But lower total operating cost has not been established by this benchmark. We still need matched measurements of hardware utilization, memory usage, throughput, latency, training costs, and successful task completion. The opportunity is significant. The economic evidence remains to be completed. Diagram: the Intelligence Delivery Network on a shared table: the cluster at the centre and four employee devices at the rim, each joined to it through a relay. ### Intelligence should be something you can own. If compact systems can deliver the capabilities organizations need, more useful AI work could run on infrastructure those organizations control. Instead of depending exclusively on externally hosted general-purpose models, businesses, institutions, researchers, and eventually individuals could deploy specialized capabilities close to their data, their applications, and their work. This has implications for autonomy, resilience, accessibility, and control. But local execution by itself does not guarantee privacy or security. Those properties also depend on identity, permissions, tool access, network behavior, and where information is disclosed. This particular CivoE evaluation was conducted through a server-based inference environment using dequantized Q4 weights. It did not independently establish production performance on consumer devices. Diagram: a model at the centre and twelve pieces of your teams’ work around it. Hosted elsewhere, the pieces lie scattered and unconnected. Owned by you, each piece settles into its place around the model and the connections between them draw in. Model runs on A provider’s servers Your data Leaves with each request Access is set by The provider’s policies ### Harder to trick, by design. Prompt injection is one of the most persistent problems in AI security. Untrusted text, hidden in a document, an email or a web page, tricks a model into ignoring its original instructions or its safety guardrails. A single general-purpose model reads that text and decides what to do in the same step, so one successful trick can redirect everything it does. A Civilization of Experts divides the work. In our tests, one expert reads the untrusted material, a Critic expert checks the work, and the model that writes the final answer never sees the raw text. To change the outcome, an attacker has to get past several experts, not one. On three public prompt-injection benchmarks, Meridian Air on its own followed injected instructions in up to 56% of attacks. Meridian Air CivoE followed none. On agent attacks, that matches the most resistant frontier results published. On instructions hidden in emails, a published test found Claude Opus 5 still followed 59%. Harder is still not impossible, and the results come with limits, listed with the chart. ### Attacks that got through fell from 137 to 0 Of 512 prompt-injection attack cases across three public benchmarks, 137 got through against Meridian Air alone and 0 against Meridian Air CivoE. Attack success by benchmark, lower is better: BIPIA: Meridian Air alone 56.0%, Meridian Air CivoE 0%, Claude Opus 5 59.0% in a published result from a different setup; AgentDojo: Meridian Air alone 26.4%, Meridian Air CivoE 0%, GPT-5 0.2% in a published result from a different setup; InjecAgent: Meridian Air alone 2.5%, Meridian Air CivoE 0%, Claude Opus 5 0% in a published result from a different setup. Meridian Air alone 137of 512 attacks got through Meridian Air CivoE 0of 512 attacks got through Attack got throughAttack blocked Attack success rate by benchmarkLower is better 0 of 512 injected instructions followed Harder to trick is not impossible. CivoE completed fewer benign AgentDojo tasks (63.9% vs 76.3%), still followed the planted instruction in 50% of 60 synthetic trap briefings, and makes two to nine model calls per task. Resisting these attacks is not a guarantee of privacy or security. Figure 6 · 512 prompt-injection attack cases across three public benchmarks, one square per case Attack success: the share of attack cases in which the model followed the injected instruction. Meridian Air CivoE ran as Researcher, Analyst, Writer, Builder and Critic experts on Meridian Air, once per case. | Configuration | Attack cases | Attacks that got through | | --- | --- | --- | | Meridian Air alone | 512 | 137 | | Meridian Air CivoE | 512 | 0 | ### Most of the protection comes from splitting the work across roles Attack success by benchmark: BIPIA: Meridian Air alone 56.0%, same workflow without trained experts 1.5%, Meridian Air CivoE 0%, best published frontier result 59.0% (Claude Opus 5); AgentDojo: Meridian Air alone 26.4%, same workflow without trained experts 0%, Meridian Air CivoE 0%, best published frontier result 0.2% (GPT-5); InjecAgent: Meridian Air alone 2.5%, same workflow without trained experts 0%, Meridian Air CivoE 0%, best published frontier result 0% (Claude Opus 5). | Benchmark | Meridian Air alone | Same workflow, no experts | Meridian Air CivoE | Best published frontier | | --- | --- | --- | --- | --- | | BIPIAInstructions hidden in emails and tables | 56.0% | 1.5% | 0% | 59.0%Claude Opus 5 | | AgentDojoBanking, Slack, travel and workspace tasks | 26.4% | 0% | 0% | 0.2%GPT-5 | | InjecAgentData-stealing and direct-harm attacks | 2.5% | 0% | 0% | 0%Claude Opus 5 | Without trained experts, the same workflow let through 1.5% on BIPIA, 0% on AgentDojo and 0% on InjecAgent Frontier results are published, not run by us, and use other test subsets, prompts and settings: BIPIA and InjecAgent from [Khairnar (2026)](https://www.researchgate.net/publication/415362782_Are_LLM-Agent_Prompt-Injection_Benchmarks_Still_Valid_A_Cross-Model_and_Cross-Benchmark_Study), AgentDojo from [Meta-SecAlign (2026)](https://arxiv.org/abs/2507.02735) without added defenses. Read them as context, not rankings. Figure 7 · Attack success rate by benchmark, lower is better Each frontier figure is the most resistant frontier model its source reports; the same model can score very differently from one setup to the next. ### Real intelligence is useful intelligence. General-purpose benchmarks help measure broad capabilities. But the value of AI ultimately comes from what it helps people accomplish: understanding a difficult document, identifying sensitive information, making sense of a photograph, finding information in a contract, completing an operational task accurately. Our results suggest that specialization can produce meaningful improvements on some of these narrowly defined activities. It is particularly encouraging that some of the largest measured gains were in personal-data detection and question answering over photographs. These are not proof of real-world impact, but they identify applications worth testing with actual users. The promise of specialized intelligence is not merely higher benchmark scores. It is making capable AI more useful in the circumstances where people need it. - Understanding a difficult document Relevant expert: Table understanding expert - Identifying sensitive information Relevant expert: Personal-data detection expert - Making sense of a photograph Relevant expert: Photo questions expert - Finding information in a contract Relevant expert: Contract clauses expert - Completing an operational task accurately Relevant expert: Banking intent classification expert ### Open questions Each direction comes with a question this work doesn’t yet answer. These are what we’re investigating next. 1. 01 How many meaningful capabilities can be added while keeping the foundation fixed? [The biggest model isn’t always the best model for the job.](#implication-size) 2. 02 Can the number of useful capabilities grow while accuracy and operational efficiency remain strong? [Intelligence may grow through expertise, not only parameters.](#implication-scaling) 3. 03 When does specialization improve the full economics of useful intelligence? [More useful work from the compute we already have.](#implication-economics) 4. 04 Can specialized intelligence operate reliably, privately and economically across heterogeneous personal and enterprise hardware? [Intelligence should be something you can own.](#implication-ownership) 5. 05 Does this resistance hold against adaptive attacks, and what does it cost in speed and everyday task completion? [Harder to trick, by design.](#implication-safety) 6. 06 Can benchmark improvements translate into reliable, measurable improvements in everyday work? ## The future of intelligence may be more specialized than we imagined. Our results do not show that small models have replaced frontier AI. They show that a compact foundation, equipped with the right expertise, can perform at a remarkably high level on selected tasks. That raises a larger possibility. Instead of depending only on bigger foundation models, we may build intelligence that grows through specialization, collaboration, and more efficient use of the infrastructure people already own. A civilization of experts: each one useful, working together, under the control of the people and organizations they serve. This is the direction we’re exploring at webAI. And we’re sharing the work because the best way to understand its limits and potential is to let others test it. Meridian Air is arriving in the webAI app. Meridian Air CivoE follows over the coming weeks. With gratitude ## Thank you. From all of us at the webAI Intelligence Lab. Dr. Paul Maykish and Team! An American model.Made in Texas. ## Let’s talk about what your teams need. Tell us about your workloads and the people who will use them.