LLM System Design and Mannequin Choice – O’Reilly

Selecting the best LLM has change into a full-time job. New fashions seem virtually every day, every providing completely different capabilities, costs, and quirks, from reasoning strengths to value effectivity to code era. This competitors creates robust incentives for AI labs to carve out a distinct segment and provides new startups room to emerge, leading to a fragmented panorama the place one mannequin might excel at reasoning, one other at code, and a 3rd at value effectivity.

AI, in a single sense, is getting cheaper quicker than any earlier know-how, at the least per unit of intelligence. For instance, enter tokens for Gemini 2.5 Flash-Lite are roughly 600 instances cheaper than what OpenAI’s GPT-3 (davinci-002) value in August 2022, whereas outperforming it on each metric. On the similar time, entry to frontier capabilities can also be turning into dearer than ever. The reason being easy: we are able to now pay instantly for extra functionality, which has led to the rise of $300+ per thirty days Professional subscription tiers.

As we speak, any developer can run succesful open-weight fashions domestically for negligible marginal value utilizing instruments like Ollama. On the similar time, enterprise programs can expertise sharp value will increase, relying on the mannequin dimension (variety of parameters, reminiscent of 3 billion, 70 billion, and even within the trillions), the variety of inner processing steps, and the amount of enter information. For builders, these are central system design selections that instantly have an effect on feasibility and value construction. For finish customers, this complexity explains why a fundamental subscription differs a lot from a premium plan with greater limits on superior fashions.

The alternatives you make in these broader improvement selections additionally decide which LLM and inference settings are optimum in your use case.

At In direction of AI, we work throughout the LLM stack, constructing purposes, designing enterprise programs, and providing on-line programs (together with one on O’Reilly), customized company coaching, and LLM improvement consultancy. In our expertise, mannequin choice and system design have change into central to getting significant outcomes from these instruments. A lot of that, in flip, will depend on the place immediately’s fashions are gaining their capabilities. Whereas scale nonetheless performs a job, latest progress has come from a broader combine of things, together with training-data high quality, post-training strategies, and particularly how fashions are used at inference time.

The Shifting Foundations of Mannequin Functionality

Whereas early positive aspects in LLM efficiency tracked intently with will increase in pretraining compute, bigger datasets, greater fashions, and extra coaching steps, this strategy now yields diminishing returns.

Latest enhancements come from a broader mixture of methods. Pretraining-data high quality has change into simply as essential as amount, with higher filtering and AI-generated artificial information contributing to stronger fashions. Architectural effectivity, just like the improvements launched by DeepSeek, has began to shut the hole between dimension and functionality. And post-training methods, particularly instruction tuning and reinforcement studying from human or AI suggestions (RLHF/RLAIF), have made fashions extra aligned, controllable, and responsive in follow.

The extra basic shift, nonetheless, is going on at inference time. Since late 2024, with fashions like OpenAI’s o1, we’ve entered a brand new part the place fashions can commerce compute for reasoning on demand. Quite than relying solely on what was baked in throughout coaching, they will now “suppose tougher” at runtime, working extra inner steps, exploring different solutions, or chaining ideas earlier than responding. This opens up new functionality ceilings, but in addition introduces new value dynamics.

These different enchancment methods have led to a transparent divergence amongst AI labs and fashions, a fast enlargement in mannequin selection, and in some circumstances, an explosion in mannequin utilization prices.

The Trendy Price Explosion: How Inference Scaling Modified the Recreation

Inference-time compute scaling has launched a brand new dynamic in LLM system design: We’ve gone from a single lever mannequin dimension, to at the least 4 distinct methods to commerce value for functionality at runtime. The result’s a widening hole in inference value throughout fashions and use circumstances, typically by components of 10,000x or extra.

Bigger fashions (dimension scaling): The obvious lever is sheer mannequin dimension. Frontier LLMs, like GPT-4.5, usually constructed with combination of specialists (MoE) architectures, can have enter token prices 750 instances greater than streamlined fashions like Gemini Flash-Lite. Bigger parameter counts imply extra compute per token, particularly when a number of specialists are lively per question.

Sequence scaling (“considering tokens”): Newer “reasoning” LLMs carry out extra inner computational steps, or an extended chain of thought, earlier than producing their last reply. For instance, OpenAI’s o1 used ~30x extra compute than GPT-4o on common, and infrequently 5x extra output tokens per process. Agentic programs introduce an extra technique of sequence scaling and an additional layer of value multiplication. As these brokers suppose, plan, act, reassess, plan, act, and so forth, they usually make many LLM steps in a loop, every incurring extra value.

Parallel scaling: Right here, the system runs a number of mannequin situations on the identical process after which routinely selects the perfect output through automated strategies, reminiscent of majority voting (which assumes the most typical reply is probably going appropriate) or self-confidence scores (the place the mannequin output claiming the very best confidence in its response is taken as the perfect). The o3-pro mannequin doubtless runs 5–10x parallel situations over o3. This multiplies the price by the variety of parallel makes an attempt (with some nuance).

Enter context scaling: In RAG pipelines, the variety of retrieved chunks and their dimension instantly affect enter token prices and the LLM’s capacity to synthesize reply. Extra context can usually enhance outcomes, however this comes at the next value and potential latency. Context isn’t free; it’s one other dimension of scaling that builders should funds for.

Taken collectively, these 4 components signify a basic shift in how mannequin value scales. For builders designing programs for high-value issues, 10,000x to 1,000,000x variations in API prices to unravel an issue based mostly on architectural selections at the moment are reasonable prospects. Reasoning LLMs, though solely distinguished for about 9 months, reversed the development of declining entry prices to the perfect fashions. This transforms the choice from “Which LLM ought to I exploit?” to incorporate “How a lot reasoning do I wish to pay for?”

This shift modifications how we take into consideration choice. Selecting an LLM is not about chasing the very best benchmark rating; it’s about discovering the stability level the place functionality, latency, and value align along with your use case.

Core Mannequin Choice Standards

When selecting a mannequin we discover you will need to first clearly establish your use case and the minimal core AI capabilities and attributes wanted to ship it.

A standard first step is to check out commonplace benchmark scores (for instance LiveBench, MMLU-Professional, SWE-Bench). These benchmarks are a helpful start line, however some fashions are tuned on benchmark information, and real-world efficiency on duties which are really related to you’ll usually differ. Filtering benchmark exams and scores by your business and process class is a beneficial step right here. An LLM optimized for software program improvement would possibly carry out poorly in artistic writing or vice versa. The match between a mannequin’s coaching focus and your utility area can outweigh general-purpose benchmarks.

Leaderboards like LMArena and Synthetic Evaluation provide broader human‑choice comparisons however nonetheless don’t exchange customized real-world testing. It helps to have a set of your personal instance questions or duties at hand to check out a brand new mannequin for your self and see the way it performs. This could embrace a mixture of straightforward duties to ascertain a baseline and difficult edge circumstances the place it’s straightforward for a mannequin to make errors.

As you progress past advert hoc testing, for any severe improvement effort, customized evaluations are non-negotiable. They have to be tailor-made to your use case and the sorts of issues you clear up. That is the one technique to really know if a mannequin, or a change to your system, is genuinely enhancing issues for your customers and your particular enterprise objectives.

Listed here are some core components we contemplate:

Multimodality is rising as a significant differentiator. Fashions like GPT-4o and Gemini can deal with not simply textual content but in addition pictures, audio, and in some circumstances video, unlocking purposes that pure textual content fashions can’t help.

Context window and efficient context window utilization are additionally key: What number of tokens or paperwork can the mannequin course of and the way a lot of that marketed context window can the LLM really use successfully with out efficiency degradation relative to duties that use much less context?

Latency is very essential for interactive purposes. Generally, smaller or cheaper fashions have a tendency to reply quicker, whereas reasoning-heavy fashions introduce delays as a result of deeper inner computation.

Reasoning is the power to scale inference-time compute and carry out multistep problem-solving, planning, or deep evaluation.

Privateness and safety are sometimes key concerns right here. For instance, if you wish to preserve your mental property personal, you will need to use a mannequin that gained’t practice in your inputs, which frequently factors towards self-hosted or particular enterprise-grade API options.

Trustworthiness can also be turning into essential and might come right down to the fame and monitor report of the AI lab. A mannequin that produces erratic, biased, or reputationally damaging outputs is a legal responsibility, no matter its benchmark scores. For example, Grok has had well-publicized points with its alignment. Even when such points are supposedly fastened, it creates a lingering query of belief: How can one ensure it gained’t behave equally sooner or later?

Moreover, the data cutoff date additionally issues whether it is for use in a fast-moving discipline.

After understanding if a mannequin meets your minimal functionality, the subsequent choice is usually on optimizing trade-offs amongst value, reliability, safety, and latency. A key rule of thumb we discover helpful right here: If the reliability achieve from a dearer mannequin or extra inference time saves extra of your or your customers’ time (valued when it comes to pay) than the mannequin prices, going with the bigger mannequin is an efficient choice!

The Professionals and Cons of Open-Weight and Closed-API LLMs

The rise of more and more aggressive open-weight LLMs, reminiscent of Meta’s Llama sequence, Mistral, DeepSeek, Gemma, Qwen, and now OpenAI’s GPT-OSS has added a essential dimension to the mannequin choice panorama. Momentum behind this open ecosystem surged with the discharge of DeepSeek’s R1 reasoning mannequin, aggressive with OpenAI’s o1 however priced at roughly 30x decrease API prices. This sparked debate round effectivity versus scale and intensified the broader AI rivalry between China and the US. Reactions ranged from “OpenAI and Nvidia are out of date” to “DeepSeek’s prices have to be fabricated,” however no matter hype, the discharge was a milestone. It confirmed that architectural innovation, not simply scale, might ship frontier-level efficiency with far higher value effectivity.

This open-model offensive has continued with robust contributions from different Chinese language labs like Alibaba (Qwen), Kimi, and Tencent (Hunyuan), and has put aggressive stress on Meta after its open-weight Llama fashions fell behind. China’s latest management in open-weight LLMs has raised new safety/IP points with some US- and European-based organizations, although we notice accessing these mannequin weights and working the mannequin by yourself infrastructure doesn’t require sending information to China.

This brings us again to the professionals and cons of open weights. Whereas closed-API LLMs nonetheless lead on the frontier of functionality, the first benefit of open-weight fashions is fast and inexpensive native testing, unparalleled flexibility, and elevated information safety when run internally. Organizations may carry out full fine-tuning, adapting the mannequin’s core weights and behaviors to their particular area, language, and duties. Open fashions additionally present stability and predictability—you management the model you deploy, insulating your manufacturing programs from sudden modifications or degradations that may typically happen with unannounced updates to proprietary API-based fashions.

Public closed-model APIs from main suppliers profit from immense economies of scale and extremely optimized GPU utilization by batching requests from 1000’s of customers, an effectivity that’s troublesome for a single group to copy. This usually signifies that utilizing a closed-source API might be cheaper per inference than self-hosting an open mannequin. Safety and compliance are additionally extra nuanced than they first seem. Whereas some organizations should use self-hosted fashions to simplify compliance with laws like GDPR by maintaining information completely inside their very own perimeter, this locations the whole burden of securing the infrastructure on the inner group—a fancy and costly enterprise. Prime API suppliers additionally usually provide devoted situations, personal cloud endpoints, and contractual agreements that may assure information residency, zero-logging, and meet stringent regulatory requirements. The selection, subsequently, shouldn’t be a easy open-versus-closed binary.

The boundary between open and closed fashions can also be turning into more and more blurred. Open-weight fashions are more and more provided through API by third-party LLM inference platforms, combining the flexibleness of open fashions with the simplicity of hosted entry. This hybrid strategy usually strikes a sensible stability between management and operational complexity.

Main Closed LLMs

Under, we current some key prices and metrics for main closed-source fashions accessible through API. Many of those fashions have extra complexity and different pricing together with choices for quick modes, considering modes, context caching, and longer context.

We current the most recent LiveBench benchmark rating for every mannequin as one measure for comparability. LiveBench is a constantly up to date benchmark designed to offer a “contamination-free” analysis of enormous language fashions by commonly releasing new questions with goal, verifiable solutions. It scores fashions out of 100 on a various set of difficult duties, with a big give attention to capabilities like reasoning, coding, and information evaluation. The same LiveBench scores between GPT-4.5 and a pair of.5 Flash-Lite, regardless of 750x enter token value variation, highlights each that smaller fashions at the moment are very succesful but in addition that not all capabilities are captured in a single benchmark!

Main open-weight LLMs

Under, we additionally current key prices, the LiveBench benchmark rating, and context size for main open-weight fashions accessible through API. We evaluate hosted variations of those fashions for simple comparability. Completely different API suppliers might select to host open-weight fashions with completely different ranges of quantization, completely different context lengths, and completely different pricing, so efficiency can differ between suppliers.

Whether or not hosted or self-deployed, choosing a mannequin solely solves a part of the issue. In follow, a lot of the complexity and alternative lies in how that mannequin is used: the way it’s prompted, prolonged, fine-tuned, or embedded inside a broader workflow. These system-level selections usually have a higher affect on efficiency and value than the mannequin selection itself.

A Sensible Information to Designing an LLM System

Merely choosing the largest or latest LLM is never the optimum technique. A more practical strategy begins with a deep understanding of the developer’s toolkit: understanding which approach to use to which downside to attain the specified functionality and reliability with out pointless value. That is all a part of the fixed “march of nines” as you develop LLM programs modularly to unravel for extra reliability and functionality. There’s a have to prioritize the best wins that ship tangible worth earlier than investing in additional incremental and infrequently expensive accuracy enhancements. The truth will all the time differ on a case-by-case foundation, however here’s a fast information to navigating this course of.

Step 1: Open Versus Closed?

That is usually your first choice.

• Go along with a closed-API mannequin (e.g., from OpenAI, Google, Anthropic) if: Your precedence is accessing absolutely the state-of-the-art fashions with most simplicity.
• Go along with an open-weight mannequin (e.g., Llama, Mistral, Qwen, DeepSeek) if:
  • Knowledge safety and compliance are paramount: If you should assure that delicate information by no means leaves your personal infrastructure.
  • You want deep customization and management: In case your aim is to fine-tune a mannequin on proprietary information and to create a specialised knowledgeable that you just management utterly.

If you happen to went open, what are you able to realistically run? Your personal GPU infrastructure is a tough constraint. Assess your cluster dimension and reminiscence to find out in the event you can effectively run a big, main 1 trillion+ parameter MoE mannequin, reminiscent of Kimi K2, or if you’re higher served by a medium-size mannequin reminiscent of Gemma 3 27B or a a lot smaller mannequin like Gemma 3n that may even run on cell.

Step 2: Gauging the Want for Reasoning

Does your process require the mannequin to easily blast out a response, or does it have to suppose first?

• Reasoning: For duties that contain advanced, multistep problem-solving, brainstorming, strategic planning, intricate code era, or deep evaluation, you want a devoted reasoning mannequin reminiscent of o3, Gemini 2.5 Professional, DeepSeek R1, or Claude 4. In some circumstances these fashions can be utilized in high-reasoning mode, which inspires the mannequin to suppose for longer earlier than responding.
• No reasoning: For simple duties like easy Q&A, summarization of a single doc, information extraction, or classification, a strong reasoning mannequin is overkill.
• The center floor: For duties requiring average reasoning, reminiscent of producing a structured report from just a few information factors or performing fundamental information evaluation at scale, a “mini” reasoning mannequin, like OpenAI’s o4-mini or Gemini Flash 2.5, provides a stability of functionality and value.

Step 3: Pinpointing Key Mannequin Attributes

Past basic intelligence and reasoning, trendy LLMs are specialists. Your selection ought to be guided by the particular attributes and “superpowers” your utility wants.

• Prioritize accuracy over value for high-value duties the place errors are expensive or the place a human knowledgeable’s time is being saved. o3-pro is a standout mannequin right here and it could possibly even be used as a reality checker to meticulously verify the small print of an earlier LLM output.
• Prioritize velocity and value over accuracy: For user-facing, real-time purposes like chatbots or high-volume, low-value duties like easy information categorization, latency and value are paramount. Select a hyper-efficient “flash” or “mini” mannequin reminiscent of Gemini 2.5 Flash-Lite. Qwen3-235B fashions may also be a fantastic possibility right here however are too advanced to inference your self.
• Do you want a deep, long-context researcher? For duties that require synthesizing data from large paperwork, total codebases, or in depth authorized contracts, a mannequin with an unlimited and extremely efficient context window is essential. Gemini 2.5 Professional excels right here.
• Is multimodality important? In case your utility wants to grasp or generate pictures, course of audio in actual time, or analyze video, your selection narrows to fashions like GPT-4o or the Gemini household. For one-shot YouTube video processing, Gemini is the standout.
• Is it a code-specific process? Whereas many fashions can code, some are explicitly tuned for it. Within the open world, Codestral and Gemma do an honest job. However Claude has gained hearts and minds, at the least for now.
• Do you want dwell, agentic internet search? For answering questions on present occasions or matters past the mannequin’s data cutoff, contemplate a mannequin with a built-in, dependable internet search, reminiscent of o3.
• Do you want advanced dialogue and emotional nuance? GPT-4.5, Kimi K2, Claude Opus 4.0, or Grok 4 do a fantastic job.

Step 4: Prompting, Then RAG, Then Analysis

Earlier than you dive into extra advanced and dear improvement, all the time see how far you will get with the best methods. It is a path of escalating complexity. Mannequin selection for RAG pipelines is usually centered round latency for finish customers, however just lately extra advanced agentic RAG workflows or long-context RAG duties require reasoning fashions or longer context capabilities.

1. Immediate engineering first: Your first step is all the time to maximise the mannequin’s inherent capabilities via clear, well-structured prompting. Usually, a greater immediate with a extra succesful mannequin is all you want.
2. Transfer to retrieval-augmented era (RAG): In case your mannequin’s limitation is an absence of particular, personal, or up-to-date data, RAG is the subsequent logical step. That is the perfect strategy for decreasing hallucinations, offering solutions based mostly on proprietary paperwork, and guaranteeing responses are present. Nonetheless, RAG shouldn’t be a panacea. Its effectiveness is completely depending on the standard and freshness of your dataset, and constructing a retrieval system that constantly finds and makes use of the most related data is a big engineering problem. RAG additionally comes with many related selections, reminiscent of the amount of information to retrieve and feed into the mannequin’s context window, and simply how a lot use you make of long-context capabilities and context caching.
3. Iterate with superior RAG: To push efficiency, you will have to implement extra superior methods like hybrid search (combining key phrase and vector search), re-ranking retrieved outcomes for relevance, and question transformation.
4. Construct customized analysis: Guarantee iterations in your system design, additions of latest superior RAG methods, or updates to the most recent mannequin are all the time shifting progress ahead in your key metrics!

Step 5: Wonderful-Tune or Distill for Deep Specialization

If the mannequin’s core habits—not its data—remains to be the issue, then it’s time to think about fine-tuning. Wonderful-tuning is a big enterprise that requires a high-quality dataset, engineering effort, and computational sources. Nonetheless, it could possibly allow a smaller, cheaper open-weight mannequin to outperform a large generalist mannequin on a particular, slim process, making it a strong software for optimization and specialization.

• Wonderful-tuning is for altering habits, not including data. Use it to show a mannequin a particular talent, type, or format. For instance:
  • To reliably output information in a fancy, structured format like particular JSON or XML schemas.
  • To grasp the distinctive vocabulary and nuances of a extremely specialised area (e.g., authorized, medical).
  • Some closed-source fashions can be found for fine-tuning through API reminiscent of Gemini 2.5 Flash and varied OpenAI fashions. Bigger fashions are usually not accessible.
  • In open-weight fashions, Llama 3.3 70B and Qwen 70B are fine-tuning staples. The method is extra advanced to fine-tune an open-weight mannequin your self.
• Mannequin distillation may function a production-focused optimization step. In its easiest kind, this consists of producing artificial information from bigger fashions to create fine-tuning datasets to enhance the capabilities of smaller fashions.
• Reinforcement fine-tuning (RFT) for problem-solving accuracy
  As an alternative of simply imitating appropriate solutions, the mannequin learns by trial, error, and correction. It’s rewarded for getting solutions proper and penalized for getting them fallacious.
  • Use RFT to: Create a real “knowledgeable mannequin” that excels at advanced duties with objectively appropriate outcomes.
  • The benefit: RFT is extremely data-efficient, usually requiring just a few dozen high-quality examples to attain vital efficiency positive aspects.
  • The catch: RFT requires a dependable, automated “grader” to offer the reward sign. Designing this grader is a essential engineering problem.

Step 6: Orchestrated Workflows Versus Autonomous Brokers

The essential choice right here is how a lot freedom to grant. Autonomous brokers are additionally extra prone to want dearer reasoning fashions with higher ranges of inference scaling. Parallel inference scaling strategies with a number of brokers are additionally starting to ship nice outcomes. Small errors can accumulate and multiply throughout many successive agentic steps so the funding in a stronger extra succesful mannequin could make all of the distinction in constructing a usable product.

• Select an orchestrated workflow for predictable duties
  You design a particular, usually linear, sequence of steps, and the LLM acts as a strong element at a number of of these steps.
  • Use when: You might be automating a recognized, repeatable enterprise course of (e.g., processing a buyer help ticket, producing a month-to-month monetary abstract). The aim is reliability, predictability, and management.
  • Profit: You preserve full management over the method, guaranteeing consistency and managing prices successfully as a result of the quantity and kind of LLM calls are predefined.
• Construct hybrid pipelines: Usually, the perfect outcomes will come from combining many LLMs, open and closed, inside a pipeline.
  • This implies utilizing completely different LLMs for various levels of a workflow: a quick, low cost LLM for preliminary question routing; a specialised LLM for a particular subtask; a strong reasoning LLM for advanced planning; and maybe one other LLM for verification or refinement.
  • At In direction of AI, we regularly have 2-3 completely different LLMs from completely different firms in an LLM pipeline.
• Select an autonomous agent for open-ended issues. You give the LLM a high-level aim, a set of instruments (e.g., APIs, databases, code interpreters), and the autonomy to determine the steps to attain that aim.
  • Use when: The trail to the answer is unknown and requires dynamic problem-solving, exploration, or analysis (e.g., debugging a fancy software program problem, performing deep market evaluation, planning a multistage venture).
  • The essential danger—runaway prices: An agent that will get caught in a loop, makes poor selections, or explores inefficient paths can quickly accumulate monumental API prices. Implementing strict guardrails is essential:
    • Finances limits: Set laborious caps on the price per process.
    • Step counters: Restrict the full variety of “ideas” or “actions” an agent can take.
    • Human-in-the-loop: Require human approval for doubtlessly costly or irreversible actions.
  • Gemini 2.5 Professional and o3 are our favorite closed-API fashions for agent pipelines, whereas in open-weight fashions we like Kimi K2.

Working via these steps helps translate a obscure downside right into a concrete implementation plan, one which’s grounded in clear trade-offs and tailor-made to your wants. This structured strategy usually yields programs that aren’t solely extra succesful and dependable but in addition far more practical for particular duties than a general-purpose chatbot ever might be.

Conclusion

The open-versus-closed race provides us fast entry to robust LLMs but in addition creates complexity. Deciding on and deploying them calls for each engineering self-discipline and financial readability.

Growing within the LLM ecosystem calls for a brand new stage of engineering self-discipline and eager financial consciousness. No single LLM is a cure-all. A sensible, evolving toolkit is crucial, however understanding which software to tug out for which job is the true artwork. The problem isn’t simply choosing a mannequin from an inventory; it’s about architecting an answer. This requires a scientific strategy, shifting from high-level strategic selections about information and safety right down to the granular, technical selections of improvement and implementation.

The success of specialised “LLM wrapper” purposes like Anyscale/Cursor for coding or Perplexity for search, a few of which at the moment are valued at over $10 billion, underscores the immense worth on this tailor-made strategy. These purposes aren’t simply skinny wrappers; they’re refined programs that leverage basis LLMs however add vital worth via customized workflows, fine-tuning, information integration, and consumer expertise design.

In the end, success hinges on knowledgeable pragmatism. Builders and organizations want a pointy understanding of their downside area and a agency grasp of how value scales throughout mannequin selection, sequence and parallel reasoning, context utilization, and agentic habits. Above all, customized analysis is non-negotiable as a result of your use case, not a benchmark, is the one commonplace that actually issues.