LMRLINE MEDIA RESEARCHINDEPENDENT R&D / JAPAN

PRIVATE AI / BUILT FOR YOUR R&D

What happened on LMR’s own dev floor.

Sometimes a local LLM is
far more capable.

Did you knowthat can happen?

Room to think and to verify, for the hard work.
Dedicated GPUs, engineering knowledge and execution tools together
carry work that stalled on cloud AI to its next stage.

A description based on LMR’s own experience in operation. Results depend on the task, model and settings, and this is not a claim of superiority on all tasks.

YOUR COMPUTE. YOUR PRIORITIES.GLM-5.2 / B300 × 4 / RAG + MCP

What we measure is how far the work actually gets.

Task completion / human rework / time to verification
01 / WHAT CHANGED01

Work that had stalled
starts moving.

What caught LMR’s attention was not how well it converses, but how far real development gets. From our own operation, we feel these changes.

01 / UNBLOCK

Get past the dead end.

On work that earlier AI use could not resolve, investigation, fixes and verification move forward.

LMR / FIELD OBSERVATION
02 / FIND

Notice what was missed.

In crawler development it surfaces points we used to miss, such as gaps in collection or processing.

LMR / FIELD OBSERVATION
03 / REMEMBER

Fill in from short instructions.

By searching past conversations and know-how from other work, it can supply conditions that were never spelled out.

LMR / FIELD OBSERVATION

Observations based on the owner’s operational reports. Not a controlled comparison that isolates cause, nor a measured performance multiple.

02 / THE ECONOMICS OF COMPUTE02

Compute you share.
Compute you own.

A flat-rate shared service must balance compute cost, latency and the throughput of all its users. With your own equipment, you decide the compute and time given to one important piece of work.

SHARED SERVICE

A typical shared cloud

User AUser BUser C
A shared inference base and quota

Compute is allocated with the demand and cost of all users in view.

DEDICATED INFRASTRUCTURE

LMR’s local environment

Priority workOngoing studyRetesting
GLM on four B300s / allocated in-house

Inference, exploration and verification continue according to your priorities.

What to compare is the sustained
budget of compute, trials and verification a task can draw on.

A consideration and conceptual model drawn from business structure. Each provider’s GPU allocation is undisclosed, and an average per user is not the same as the GPU scale used for a single inference. Dedicated contracts differ in their terms. [R4]

03 / WHAT PROVIDERS DISCLOSE03

Tuning the amount of reasoning
is officially documented.

Even for the same model, there are settings that tune speed, cost and the depth of the response. Maximum reasoning and verification on every task is not necessarily the default.

OPENAI / REASONING EFFORT

Tune reasoning depth
against speed and token use.

The official API docs explain that a low effort setting favors speed and fewer tokens. Higher reasoning settings can also be chosen.

Read the official docs ↗

ANTHROPIC / EFFORT

At low effort,
fewer tool calls too.

The official API docs describe trading cost and speed against thoroughness, and note the effect of low effort on capability and a reduction in tool calls.

Read the official docs ↗

Official docs reviewed on 16 November 2026. These are published API specifications; they do not prove any app’s internal settings or the cause of a particular failure. They show the trade-off between compute and quality.

04 / SUSTAINED WORK04

Stop partway,
and solvable problems remain.

Try a hypothesis. Read the failure. Look for another way. Fix it, and go through to regression tests.
On hard work, this iteration decides the outcome.

INPUT

Reasoning & execution budget

Time, tokens, iteration limits

→
EXPLORE

Explore & try

Investigate alternatives, run tools

→
VERIFY

Where verification lands

Fix, retest, completion check

When it falls short

Only candidate causes offered
partial fix, left unverified

When iteration can continue

Alternatives verified, fixes with evidence
on to retest and completion

On hard work, the choice to spend the compute.

“Look again,” “try another way,” “check after fixing.” Dedicated GPUs combined with execution and evaluation.

A shortage of compute or execution budget can be one cause of unsolved work. But model capability, instructions, knowledge, tools and evaluation method also shape the result. Long reasoning or large equipment alone do not guarantee better performance. [R1–R3]

05 / THE WHOLE SYSTEM05

Performance is set by
more than the model name.

Think, look things up, remember, try and fix.
Run this whole chain of work in one environment.

PRACTICAL PERFORMANCE

Model × compute environment
× knowledge × execution & verification

A concept for thinking about practical performance.
01

Dedicated compute

Allocate inference time and parallelism by your own priorities.

02

Accumulated know-how

Past decisions, failures and reasons for fixes, referenced via RAG.

03

Connected execution tools

Access to code, the web, images and tests.

04

A repeating improvement loop

Check the results and feed them into the next fix.

06 / FREEDOM TO BUILD06

Not being jerked around by limits
in R&D is a relief.

How far to push your own research.
Keep that decision, and the environment to act on it, in your own hands.

Do not stall the inquiry

Your research plan is less at the mercy of blanket limits or quotas on external AI services.

Choose the conditions

Design the model, tools, execution permissions and update timing around your research goals.

Try as often as you like

No per-token metering on local inference. Plan the budget for iteration within equipment, power and operating costs.

This freedom means deciding operating policy for yourself. Model and software specifications and legal or contractual terms still apply. If you also use an external API, its fees apply separately.

07 / DEVELOPMENT POSSIBILITIES07

The thing you wanted to build,
running.

Web collection, maintenance work, image understanding, continuous improvement — and AI-driven penetration testing. From one shared AI base to a system for each job.

PENTEST / DEVELOPMENT EXAMPLE

AI-driven penetration testing.

On your own systems, or an environment you were asked to assess, it links test planning, execution, evidence and fix support. Form a hypothesis, verify, and check once more after the fix.

An example of a system we could build. The features of each theme are designed to your requirements.

WORKFLOW / CONCEPT

  1. SCOPEAgreed targets, time and test scope
  2. PLAN / RUNTest plan and execution within approved limits
  3. EVIDENCELogs, impact and reproduction conditions
  4. FIX / RETESTFix support and retesting
What you take away

Test records / evidence-backed report / fix proposals / retest results

Switch themes to see the development flow and deliverables.Discuss this build ↗
08 / HARDWARE ARCHITECTURE08

A dedicated compute environment
at the center of R&D.

A configuration based on LMR’s equipment. For your site, it is designed around the target model and the number of concurrent users.

DEVELOP

Developer workstation

Launcher / IDE / CLI

CONNECT

Internal network

Authentication & access control
management / execution traffic

Request→

DELL / POWEREDGE XE9780

NVIDIA B300 × 8

GPU 01GPU 02GPU 03GPU 04GPU 05GPU 06GPU 07GPU 08

Four assigned to a local GLM-5.2

The use of the remaining four is not specified in this diagram.

400Gbps→
STORE / CONNECTX-7

200 TB SSD

Code / models / RAG
logs / verification artifacts

EXTEND

Auxiliary & expansion

VLM / search / test execution
placement set by the workload

Other equipment LMR holds

A100 × 8 / A6000 × 2 / Mac Studio M3 Ultra 512GB × 4

A schematic based on the equipment described in the source document. It does not show actual wiring or the placement of all software, and it is not a statement of any partnership, certification or sales agreement with Dell or others.

09 / SOFTWARE ARCHITECTURE09

From the usual workflow,
into in-house knowledge and tools.

Connect the model, the execution environment and where knowledge is kept. Search for what is needed, confirm it with tools, and feed the result into the next inference.

01 / WORKSPACE

Launcher

Pick a project and settings

Claude Code

Planning / tool calls / checking results
the entry point for work today

02 / LOCAL INTELLIGENCE

Local inference API

GLM-5.2 / four B300s

MCP & tool layer

Calls the functions needed and
returns observations to the next inference

03 / KNOWLEDGE & TOOLS

RAG DB

Past conversations / know-how / evidence

VLM & in-house crawler

Understanding images and screens / web collection
the VLM connects via MCP

Integrated when productized

Git / CI & test environments / job management / permissions & approval / monitoring & audit logs / registering verified knowledge into RAG

A logical configuration based on current operating elements. The API translation layer, the RAG search path and the compatibility and terms of third-party software are confirmed per configuration, and the execution layer is designed to be replaceable.

10 / KNOWLEDGE THAT WORKS10

The instincts of a skilled user,
made a team asset.

What to check, where it failed, why that design. Knowledge drawn from past conversations, delivered in a form usable in today’s work. Even without updating model weights, the experience it draws on can be grown.

01 / CAPTURE

Conversation & run logs

Instructions, reasons for fixes
the story of successes and failures

02 / ORGANIZE

Organize into knowledge

Conditions and evidence
confidentiality class and expiry

03 / RETRIEVE

Search when needed

Narrow to the knowledge
that fits the task at hand

04 / VERIFY

Run and verify

Confirm the result and
update the usable knowledge

Verification updates the knowledge ↩

Not just stuffing in knowledge, but designing which task it serves, and when.

11 / CONTINUOUS IMPROVEMENT11

Set a goal, then
run the improvement cycle.

With round-the-clock operation in view, trial, evaluation and re-execution are automated. Before telling the AI to “just make it better,” we build the criteria to confirm it actually did.

If improvement stalls, it records why and hands off to a person.

Design example / reducing crawler collection gaps

Collection rates are compared on a fixed set of test URLs. Per-site load limits and regression tests for existing features are maintained.

Goal, budget, stop conditions

01

Form a hypothesis

Study the failure logs

02

Make the fix

Implement a candidate

04

Decide the next step

Adopt / retry / stop

03

Evaluate

Check with fixed tests

01 → 02 → 03 → 04 → next hypothesis

Promotion to production follows the approval conditions you set.
12 / VERIFY THE DIFFERENCE12

Prove the difference
on your own work.

Compare not just by model name but with knowledge, tools and work budget matched. A task that is stuck now becomes the starting point for evaluation.

A / BASELINE

Your current cloud setup

Record how far it gets with your usual model, settings and quota.

B / MORE BUDGET

Cloud with more budget

With the same model, widen the reasoning, output and execution budget and see whether it improves.

C / LMR LOCAL

LMR’s local environment

On the same task, RAG and tool conditions, measure completeness and the effort required.

What we evaluate

Completion rate / undiscovered defects / regression tests / human rework / time taken / total cost

A proposed comparison plan. B is evaluated within what the service allows you to change. Because different models also bring differences in model capability, the effect of RAG and compute budget is checked under separated conditions.

13 / FROM OUR LAB TO YOUR TEAM13

Honed as a user,
delivered ready to use.

LMR is a lab with its own compute equipment and development work. The operational instincts gained there carry into your work. Equipment, software and knowledge base are designed as one development environment.

  1. 01

    Compare on real work

    Agree the task to solve and the pass conditions.

  2. 02

    Connect environment and knowledge

    Combine GPUs, model, RAG and tools.

  3. 03

    Operate as a team

    Set up usage, maintenance, updates and evaluation.

LMR in-house AI compute
LMR / IN-HOUSE AI INFRASTRUCTURE
Photo of our own equipment / from lmr.jp
14 / QUESTIONS

Before adopting,
what you will want to check.

Is a local LLM always more capable than the cloud?+

Performance varies with the task, model, reasoning settings, RAG, tools and execution budget. The wording on this page is based on LMR’s own experience in operation. On your real work, we compare completion rate and the effort required.

Are there really no token fees?+

Inference on a local model you run yourself has no per-token cloud-API metering. Equipment, power, maintenance and operating costs do apply. If you combine an external API or another paid service, each has its own fees.

Can past conversations and in-house know-how be used?+

We consider a configuration that organizes conditions, evidence and confidentiality class, and searches RAG for knowledge that fits the task at hand. Managing the freshness and quality of that knowledge, and updating it from execution and verification, is key.

Can we keep our current workstations and dev tools?+

We design a configuration that connects your launcher, IDE or CLI to the local inference API. Model-connection compatibility and each tool’s terms are checked individually, and the execution layer is reselected as needed.

How much GPU equipment is needed?+

It is designed from the target model, quantization, context length, number of concurrent users, response speed and the continuous workloads to run. The B300 configuration shown is LMR’s own example and does not require the same equipment of every customer.

LET’S BUILD

Your next build,
on your own compute.

Bring us one piece of work that is stuck right now.
How far local AI can carry it forward —
we check it together, against the real task.

What we ask in a consultation: what you want to build / where you are stuck / number of users / data to keep in-house / existing compute equipment

About this page and its references +

This page is based on “LMR Local AI Platform, web-page draft (revised)” (16 November 2026). It distinguishes operational experience, published specifications, conditional considerations and development proposals.

  1. R1OpenAI / Reasoning models ↗
  2. R2Anthropic / Effort ↗
  3. R3OpenAI / Latency optimization ↗
  4. R4NVIDIA / Triton Inference Server — Batchers ↗

Official docs reviewed on 16 November 2026. They explain the technical background of shared inference and the tuning of reasoning volume; they are not proof of any specific provider’s GPU allocation, nor of a practice of deliberately leaving tasks unfinished.