In Ken Imoto’s September 29, 2026 experience report, a local model was judged unsuitable for 96 of 100 whole requests—but 97 of 200 individual agent steps were judged suitable for local execution. Those are separate counts, not a single success rate. The distinction points to a practical hybrid: keep difficult planning on a frontier model and send only well-defined, lower-risk steps to a local model. Imoto’s results describe his own setup and sample, not a general benchmark of local LLMs or Claude Code.
What Imoto tested—and what his numbers mean
Imoto reports running qwen3.5:4b locally on an RTX 4070 while considering whether to use Ollama with Claude Code. His configuration is an example, not a hardware recommendation or minimum requirement. In his sample, he judged 96 of 100 whole requests to require a frontier model. Separately, he judged 97 of 200 individual agent steps suitable for local execution. The request count and step count use different units and denominators; one cannot be converted into the other.
As an Amazon Associate I earn from qualifying purchases.
The underlying idea is that a short request can still require difficult planning, while a discrete operation inside that request may be straightforward. A fetched page, for example, can require a bounded extraction step even if deciding what to research and how to act on the result calls for stronger reasoning. Imoto reports 20 real WebFetch examples; extraction was imperfect, with some pages failing to load and some answers partly wrong. His results are self-reported, and the source does not establish independent replication across other models, hardware, task mixes, or Claude Code versions. Read Imoto’s account.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallWhy route steps instead of entire requests?
A whole-request decision asks whether a model can understand the goal, plan the work, handle ambiguity, and recover when something goes wrong. A step-level decision asks whether one operation has a narrow input and a checkable output. Those are different judgments. A request that needs a frontier model for planning may still contain a fetch, formatting, or short summarization step that a local model can attempt.
#1 Best Overall
That makes hybrid routing a workflow-design choice, not proof that local execution is broadly capable. Start from steps that are bounded, observable, and easy to reject or retry. Track actual outcomes in your own workflow: success, incorrect output, tool failure, retry, and escalation. Imoto’s counts can motivate that measurement, but they do not predict yours.
What a conservative hybrid route looks like
Imoto’s example uses explicit rules rather than trusting a probability router to choose every model. In that configuration, WebFetch is sent local, while Agent and Task stay on the frontier model. Commit-message generation also remains on the frontier in the shown example, probability routing is off by default, and the default route is frontier. He says one explicit WebFetch rule performed better than an earlier 9B probability router in the case he tested; that is an observation from his setup, not a universal result.
| Work or route | Imoto’s example | Practical implication |
|---|---|---|
| Default | Frontier model | Unclassified or unexpected work does not silently move to the local model. |
WebFetch |
Local model | A narrow page-processing step is explicitly assigned locally; failed or unreliable output still needs handling. |
Agent and Task |
Frontier model | Broader agent work remains on the stronger route in this example. |
| Commit-message generation | Frontier model | The shown configuration does not move this task to local execution. |
| Probability routing | Off by default | Routing is based on explicit rules in the example rather than a probabilistic classifier. |
Imoto describes Claude Code’s WebFetch tool as using a small, fast model for fetched-page processing. That is his description of the product behavior he encountered; it should not be taken as confirmation of the current implementation for every release. Check current Claude Code documentation before relying on product details.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →How to decide whether a step belongs locally
Use a route only when its boundaries and fallback behavior are clear. Before moving a step off the frontier route, answer these questions:
- Is the task narrow? Prefer a defined transformation or extraction over open-ended planning.
- Can you validate the result? Decide what constitutes a usable answer and reject output that fails the check.
- What happens on failure? Specify whether a failed fetch, malformed response, or uncertain result triggers a retry, frontier fallback, or human review.
- Can the data go to this model? Set the route according to your data-handling requirements, not just token cost.
- What is the full workflow cost? Include local runtime, frontier calls, retries, and the context the orchestrator must read back.
- Does it work on your hardware and workload? Imoto’s RTX 4070 and model are one reported setup; his account does not establish a general hardware floor or comparative throughput.
Keep the default conservative while collecting evidence. If local output fails, is incomplete, or cannot be checked, route that case to a stronger model or a person rather than treating an attempted local run as a successful one.
Why a local model should not be your only secret detector
Imoto reports a serious false negative in his local judge: it approved a note containing a production database password, giving it 69% confidence. In his labeled examples, a 0.5 threshold missed 20 secrets; a threshold selected using those same examples missed 2. He also says six real secrets did not match the evaluation set discussed in his longer write-up. These are results from his judge and data, not validated security rates for other systems. A confidence threshold—even one tuned on examples—is not a security guarantee.
Rank #4
Do not rely on a local LLM as the sole secret detector or approval gate. Use an independent, tested secret-scanning layer where appropriate, and make sure sensitive data is handled according to your organization’s rules. Treat model-based classification as an additional signal, with a safe failure path, not as proof that content is clean.
“Free” local tokens do not guarantee a cheaper workflow
Imoto says his local-worker arrangement was the most expensive configuration he tested because the orchestrator reread the worker’s output. As he puts it: “The local model’s tokens were free. The orchestrator re-reading everything the worker sent back was not.” This is a reported outcome from his setup, not a cost comparison that applies to every workflow.
Best Value
Measure the full path rather than comparing only local token charges with frontier-model usage. Imoto’s example constrains the worker’s return value to typed fields, limits the summary to two lines, and writes full logs to a file. The broader design lesson is to return only what the next step needs and keep bulky detail out of repeated orchestration context. Whether that saves money in another setup depends on its actual local runtime, model calls, retries, and context handling.
Implementation details that can change
Imoto describes using Ollama’s native /api/chat endpoint with log probabilities for a constrained Y/N judge. In his setup, the OpenAI-compatible endpoint dropped probabilities. He also warns that thinking models may not put the answer in the first token, and that candidate tokens outside the returned top-logprob set can appear to have zero probability. These details matter if a router or judge depends on token probabilities: an endpoint or output-format assumption can invalidate the decision logic.
The article says its companion scripts expect Ollama 0.12.11 or later. That is a version requirement for those scripts as described in the post, not a claim that this is the current Ollama release or a recommendation to install that version. Verify current Ollama documentation and the project’s own instructions before adapting the implementation. The source also identifies companion local-step-router code as MIT licensed; its current state and compatibility are not established here.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
