The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
A local MCP codebase-memory server can keep project indexing and answer synthesis on your own machine, but its usefulness depends on two separate things: retrieving relevant code and controlling how many inference requests the GPU handles at once. In a July 15, 2026 post-mortem, product engineer Enrique Bruzual describes how the zerikai_memory project combines ChromaDB retrieval with Ollama synthesis—and how unbounded concurrent requests overwhelmed an 8 GB GPU.
What the local MCP server does
Bruzual describes zerikai_memory as supporting cloud, local, and hybrid modes. In local mode, Ollama performs synthesis while ChromaDB supplies structured project entities, including function signatures, file paths, line ranges, and docstrings. The system combines those entities with a project brief and returns an answer with inline #file:line citations. These are details of the project described in Bruzual’s July 15, 2026 post, not a universal MCP specification.
The architecture separates retrieval from synthesis. Bruzual says retrieval uses ChromaDB with L2 distance and lexical reranking, so the synthesis model can be changed without changing the retrieved context. That separation makes a fairer model comparison possible: give both models the same entities and system prompt, then compare their answers.
The account is a report on this project’s implementation and tests, including its “universal-brain” MCP layer, rather than a step-by-step recipe covering every MCP client or deployment. A separate codebase semantic-search MCP repository documents stdio transport and Ollama embeddings; it is an independent example, not part of zerikai_memory. ChromaDB’s Go client documentation also lists Ollama as an embedding integration, which establishes that the combination is available, not that every implementation uses it in the same way.
What the model comparison shows—and does not show
Bruzual’s standalone latency test sent static ChromaDB payload samples, built from real workspace entities, to the models. It measured raw latency at the HTTP layer using three queries and three samples per model. The author explicitly distinguishes this from measuring synthesis quality on a live codebase.
| Model | Mean latency | Standard deviation | Minimum–maximum |
|---|---|---|---|
mistral:7b |
6.14 seconds | 3.58 seconds | 2.92–14.57 seconds |
ornith:9b |
13.39 seconds | 5.76 seconds | 8.77–25.67 seconds |
These are Bruzual’s 2026 measurements on one Windows 11 computer, not independent benchmarks or performance guarantees. The first ornith:9b request took 25.67 seconds as memory spilled into shared system memory before Ollama pinned the model; warmed samples were reported at 9–17 seconds. Bruzual reports that mistral:7b fit within the tested 8 GB of dedicated VRAM and ran in 3–7 seconds when warm. Cold-start behavior and GPU memory pressure can therefore matter as much as the model name.
Evaluate grounding, not just speed
In a separate live test, Bruzual sent five queries through the project’s MCP layer and supplied both models the same retrieved context and system prompt. One failure case exposed a key distinction: when retrieval did not provide enough context, one model gave a confident answer with unsupported details while the other said it could not determine the answer. A citation format alone does not prove an answer is grounded; check whether cited files and lines actually support its claims, and whether the model recognizes missing evidence.
Why the brief-generation comparison is inconclusive
Bruzual also compared a cloud DeepSeek run with local ornith for brief generation, but the comparison was uncontrolled because docstring density differed between runs. It does not establish that either model is generally better. The author’s interpretation is that richer indexed context improved the generated brief: “The takeaway is not that ornith beats DeepSeek for brief generation. It is that embedding-docstring enrichment is visible and measurable in the output.” The practical implication is to control index content before attributing an output difference to the model.
Rank #3
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
Why local synthesis saturated the GPU
The tested machine was a Windows 11 system with an NVIDIA RTX 3050 with 8 GB of dedicated GDDR6 VRAM, an Intel i7-12700 CPU, and 32 GB of RAM. Bruzual reports that the local deep-brief workflow launched all nine brief-section tasks with asyncio.gather and no concurrency gate. That allowed several Ollama requests to compete for the limited GPU memory and saturated the card.
The reported fix was to add a global ollama_semaphore and a safe wrapper. In local mode, the wrapper uses the semaphore to limit Ollama requests; cloud and hybrid calls can bypass it. The project exposes OLLAMA_MAX_CONCURRENCY, with a reported default of 1 for 8 GB hardware. This is the author’s configuration for that workload, not a universal setting for every GPU or model.
Rank #4
- 【AI Max+ 395 AI Workstation】16 cores, 32 threads, up to 5.1 GHz boost and 80 MB cache. Integrated Radeon 8060S graphics with 40 CUs, RDNA 3.5, delivers performance close to RTX 4060/4070 laptop GPUs. Triple-engine design(CPU+GPU+XDNA 2 NPU) with up to 126 TOPS total, including 50+ TOPS dedicated NPU for local AI inference and machine learning acceleration. Ideal for AI development, content creation, virtualization, data analysis, and demanding multitasking. Compact, high-performance workstation.
- 【256-bit LPDDR5X MAX 128GB】The LPDDR5X onboard memory reaches 8400 MT/s - 1.5x faster than DDR5 SODIMM. Unlock the full potential of your graphics with massive 128GB memory pooling. This system allows you to manually assign up to 128GB of the onboard RAM to serve as video memory (VRAM) directly within the BIOS setup, delivering unparalleled performance for 4K video editing, and AI model training without the need for a discrete graphics card.
- 【Lastest GPU 8060S & XDNA 2 NPU】Built on the RDNA 3.5 architecture, the AMD Radeon 8060S Graphics iGPU features 40 compute units (2,560 stream processors). It delivers performance on par with NVIDIA's mobile RTX 4070, efficient encoding/decoding for AVC, HEVC, VP9, and AV1 video codecs. And It can connect 4 screens via HDMI & DisplayPort & Full Featured USB4 x2 to efficiently handle your tasks and meet your specific needs. Supports 8K/4K resolution displays.
- 【Dual LAN (2.5GbE+10GbE)& WiFi 7】The computer has double LAN, one is 2.5GbE (I226), the other is 10GbE(AQC113). provides more applications, such as firewall, soft routing, multichannel aggregation. Built-in WiFi module, support WiFi 7 and Bluetooth5.4. Known as 802.11be, Wi-Fi 7 promises up to 46Gbps theoretical throughput, making it 4.8x faster than Wi-Fi 6. and computer has 4 built-in NVMe SSD slots, 1 SD card slot, allowing you to expand its storage capacity.
- 【Engineered to Endure】The computer measures 7.13 x 7.24 x 2.99 inches. AI mini pc is encased in a premium all-aluminium chassis. Dual turbo CPU fans deliver silent, ultra-efficient cooling, To enable the computer to maintain stable operation for a long time. We offer up to 2 years warranty and lifetime professional customer service. Please feel free to contact us if any issues happened. thanks
The general engineering lesson is to treat local inference concurrency as a resource limit, not as an automatic benefit of asynchronous code. Increase the limit only after observing memory use and response behavior on the actual workload; concurrent requests can increase contention rather than throughput when VRAM is scarce.
How much VRAM this workflow needs
For the exact setup he tested, Bruzual describes 8 GB of dedicated VRAM as the limit, with shared system memory over PCIe slow enough to affect inference. He recommends 10–12 GB of dedicated VRAM for more headroom in this particular workflow and names an RTX 3060 12 GB as an example. This is a setup-specific recommendation, not a requirement for all MCP servers, Ollama use, or code search.
Best Value
- 【AI Max+ 395 AI Workstation】16 cores, 32 threads, up to 5.1 GHz boost and 80 MB cache. Integrated Radeon 8060S graphics with 40 CUs, RDNA 3.5, delivers performance close to RTX 4060/4070 laptop GPUs. Triple-engine design(CPU+GPU+XDNA 2 NPU) with up to 126 TOPS total, including 50+ TOPS dedicated NPU for local AI inference and machine learning acceleration. Ideal for AI development, content creation, virtualization, data analysis, and demanding multitasking. Compact, high-performance workstation.
- 【256-bit LPDDR5X MAX 128GB】The LPDDR5X onboard memory reaches 8400 MT/s - 1.5x faster than DDR5 SODIMM. Unlock the full potential of your graphics with massive 128GB memory pooling. This system allows you to manually assign up to 128GB of the onboard RAM to serve as video memory (VRAM) directly within the BIOS setup, delivering unparalleled performance for 4K video editing, and AI model training without the need for a discrete graphics card.
- 【Lastest GPU 8060S & XDNA 2 NPU】Built on the RDNA 3.5 architecture, the AMD Radeon 8060S Graphics iGPU features 40 compute units (2,560 stream processors). It delivers performance on par with NVIDIA's mobile RTX 4070, efficient encoding/decoding for AVC, HEVC, VP9, and AV1 video codecs. And It can connect 4 screens via HDMI & DisplayPort & Full Featured USB4 x2 to efficiently handle your tasks and meet your specific needs. Supports 8K/4K resolution displays.
- 【Dual LAN (2.5GbE+10GbE)& WiFi 7】The computer has double LAN, one is 2.5GbE (I226), the other is 10GbE(AQC113). provides more applications, such as firewall, soft routing, multichannel aggregation. Built-in WiFi module, support WiFi 7 and Bluetooth5.4. Known as 802.11be, Wi-Fi 7 promises up to 46Gbps theoretical throughput, making it 4.8x faster than Wi-Fi 6. and computer has 4 built-in NVMe SSD slots, 1 SD card slot, allowing you to expand its storage capacity.
- 【Engineered to Endure】The computer measures 7.13 x 7.24 x 2.99 inches. AI mini pc is encased in a premium all-aluminium chassis. Dual turbo CPU fans deliver silent, ultra-efficient cooling, To enable the computer to maintain stable operation for a long time. We offer up to 2 years warranty and lifetime professional customer service. Please feel free to contact us if any issues happened. thanks
For hardware below 8 GB VRAM—or when latency matters more than citation precision—Bruzual recommends mistral:7b based on its faster results on his tested system. Readers should treat that as a starting point rather than a model ranking: the post’s small sample does not settle performance on other GPUs, codebases, or index contents.
A practical way to compare models on your codebase
- Hold retrieval constant. Use the same query, retrieved entities, project brief, and system prompt for each model. Otherwise, a change in indexed context may be mistaken for a model improvement.
- Measure cold and warm latency. Record first-request behavior separately from subsequent requests, since model loading and memory movement affected the reported results.
- Check citation accuracy. Verify that each file and line citation supports the claim it accompanies; do not count citation presence alone as proof of grounding.
- Test missing-context behavior. Include questions whose answers are absent from the retrieved entities and note whether the model acknowledges uncertainty or invents details.
- Observe VRAM under the real workload. Test the concurrency setting used by your application, not just a single isolated request, and constrain local requests if the GPU becomes saturated.
- Keep index quality comparable. Include equivalent docstring and source coverage before drawing conclusions about synthesis quality.
This evaluates the dimensions the post-mortem leaves open for other machines: latency, grounding, abstention, memory use under concurrency, and sensitivity to index density. Its benchmark is useful as a concrete report from one system, not as a substitute for a test on your own project.
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.
Free tools Windows power users keep installed
One-click scans. No signup required.
