Personal knowledge base · Agents · Model and hardware data
myce is a local-first knowledge base app. Material goes in, gets organized, and each library turns into an agent that talks about that material. The index and the source text stay on your own machine.
The news, rankings and hardware figures here are for one question: which model, on which machine. Every number links to its source.
Or try it in the browser, no download →
Last updated 2026-11-02 · Figures were retrieved from the linked sources on that date
Product
Windows v1.2.0 · iOS 1.0.2
The name comes from mycelium. Underground, mycelium joins scattered roots into one network and moves nutrients through it.
myce does the same with material: pulls what is scattered into one library, then lets that library speak.
Runs local models through LM Studio or Ollama, or points at a cloud API. When an answer is worth keeping, send it straight into a library.
Local model · cloud API
Create a library, import a whole folder, export the lot as one .myce-kb file. Index and source text stay on your machine.
.myce-kb
An agent grows out of a library. Swap the library and you get a different partner, one that knows only that subject.
Built from a library
Send a .myce-kb to someone, or import one they built. Same format in both directions.
Import · export
The CLI calls each module directly and returns JSON, so it drops into scripts or another agent framework.
JSON out
Models and data stay local and work offline. Reach for a cloud model only when you want more horsepower.
Works offline
Download
Windows v1.2.0 · portable
Unzip and run; there is no runtime to install. On first launch you pick a model: LM Studio or Ollama on the same machine, or a cloud API endpoint.
Download for WindowsInstall it on iPhone and iPad, free, with no separate backend to configure. Listed in the China App Store.
Get it on the App StoreThe CLI ships inside the desktop build. Call modules directly, get JSON back, wire it into your own scripts.
To share one library across devices, run the service in Docker: library upload and download, agent chat and knowledge mirrors all live there.
News
9 articles
A prototype from DeepSeek's Harness team watches developer group conversations, spots coding problems and proposes fixes without waiting for a prompt. The code is open source. The announcement came with no benchmarks and no stated license.
It scores 59 on Intelligence Index v5.0, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Access starts with selected cybersecurity defenders through the Fairwind Program, with general release expected in a few weeks. Introductory pricing is $2 input and $10 output per million tokens.
On its Q3 earnings call, DoorDash said third-party AI agents have sent it only several hundred orders, a tiny share of its daily volume. It is an early data point on how much shopping agents actually do.
Zest finds restaurants and books reservations by text message, uses Plaid to read card transactions for recommendations, and lets groups vote in the thread. Separately, the privacy-focused assistant app Ollie is closing after failing to raise enough capital or win enough users.
The multimodal mixture-of-experts model has three reasoning modes. Mistral reports 48% on the Artificial Analysis Coding Agent Index, roughly level with GLM-5.3, while using 47% fewer tokens per task. Those are Mistral's own figures, and the report advises checking cost on your own workloads.
Fusion routes each request to a model or a combination of models and picks the reasoning effort, so you choose an effort level instead of a model. It is on all Copilot plans with no separate fee, and paid plans get a temporary 30% discount on model costs. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the cost per task. Those are GitHub's own figures.
The M5 Ultra configuration with 512GB of unified memory and 1TB of storage costs $22,299, and the memory upgrade alone adds $15,500 over the 96GB base. Orders are expected to arrive from late November to early December.
The API opened on October 23 at $2 input and $10 output per million tokens. Moonshot lists 2.8 trillion total parameters with about 104 billion active per token, and says open weights will follow by November 4.
Nvidia's 128GB model launched at $3,999 about a year earlier, and ServeTheHome ties the increase to rising memory prices. The 64GB version ships through Acer, ASUS, Dell, Gigabyte, HP and MSI from October 23, and two of them can be clustered into 128GB for around $8,000.
Rankings
Artificial Analysis Intelligence Index v5.0 · retrieved 2026-11-02 · click a column to sort
| Model |
|---|
Source: Artificial Analysis leaderboard, top 12 entries. Index v5.0 was published on 27 Oct 2026 and combines nine evaluations. Each entry is a model plus a reasoning-effort setting, shown in brackets. Cost per task is Artificial Analysis's figure in USD, and speed is the median output speed. Gemini 4 Carbon had limited access at launch. Scores are third-party and change when the index updates.
Hardware
Specs, prices and reported speeds · retrieved 2026-11-02
M5 Max goes up to 128GB of unified memory at up to 614 GB/s, from $2,499. M5 Ultra has a 30-core CPU and 64-core GPU, configurable to 36 CPU cores and 80 GPU cores; memory starts at 96GB with 256GB or 512GB available, bandwidth is 1.2 TB/s, and the 96GB base configuration starts at $5,499. A machine with 512GB of memory and 1TB of storage is $22,299, a figure you cannot reach by adding the memory premium to the base price: the memory accounts for $15,500 of it, and the rest is the storage and chip configuration. Apple says M5 Max prompt processing in LM Studio is up to 3.9x faster than M4 Max, a best-case figure from Apple's own testing.
Apple Newsroom, Aug 25 · 9to5Mac, Oct 28 · unwire.hk tested, Sep 21
128GB of unified LPDDR5x at 273 GB/s with CUDA support. The 128GB model is now $6,950, up from $3,999 at launch, and a 64GB version starts at $4,999.
ServeTheHome, Oct 3 · Spheron
Mini PCs with up to 128GB of unified memory. The benchmark machine below used LPDDR5x-8000. In March 2026, 128GB models were listed at about $2,400 to $3,200, and the source blames rising memory prices, so current prices may differ.
Liliputing, prices as of Mar 24
Fast while a model fits in VRAM. When it does not, speed drops sharply: an RTX 4080 running gpt-oss-120b with the model split 78% CPU and 22% GPU decoded at 12.45 tokens/s.
Spheron, Oct 7
| Machine, gpt-oss-120b | Decode tok/s | Prefill tok/s | Setup | Source |
|---|---|---|---|---|
| NVIDIA DGX Spark 128GB unified | 41 | 1,159 | Ollama, MXFP4 | Spheron |
| Mac Studio model not stated | 34, falling to 6 at long context | Not stated | Ollama, MXFP4 | Spheron |
| RTX 4080 plus system RAM 78% CPU / 22% GPU split | 12.45 | 969 | Ollama, MXFP4 | Spheron |
| Ryzen AI Max+ 395 128GB LPDDR5x-8000 | 52 | 404 | llama.cpp b8123, Vulkan, Q4_K_M | From the Matrix, Mar 5 |
These rows come from different tests, backends, quantizations and dates, so they are not a head-to-head comparison. Spheron notes that no published benchmark has run DGX Spark and the M5 Ultra Mac Studio side by side. The Ryzen test used a 7,624-token summarization prompt, and its author says numbers drift slightly between runs. The Mac Studio test did not state which model it used, so it cannot be read as an M5 Ultra figure.
| Model, Mac Studio M5 Ultra MTPLX runtime, test machine 256GB / 4TB | Architecture | Decode tok/s | First token | Memory | On M3 Ultra |
|---|---|---|---|---|---|
| Qwen 3.8 27B Optimized Quality | 27B dense, 8-bit | 80.3 | 0.54 s | 33.7GB | 56.4 tok/s, 1.50 s |
| Qwen 3.8 Flash-Next Optimized Speed | 125B-A6B MoE, Dynamic 4-bitsparse attention 8-bit | 114 | 0.43 s | 109.4GB | 69.4 tok/s, 1.10 s |
These are the reviewer's own measurements, not vendor figures, and they are not a head-to-head: the two rows use different models and quantizations, and neither belongs in the gpt-oss-120b table above. The task was a Python debugging problem adapted from LeetCode Hard, seeded with two logic bugs, that the model had to find and fix. Both machines ran a fresh session with a cold cache, MTP depth 3 and auto reasoning effort, and the comparison is generation speed rather than time to finish the task. Measured memory bandwidth on the same bench: 688.47 GB/s on M3 Ultra and 1,035.82 GB/s on M5 Ultra (StreamBench GPU Triad). GPU power was about 67.7W on the 27B run and 48.0W on Flash-Next, roughly 6% above M3 Ultra. Memory figures are what the reviewer's dashboard reported during the run.
Method
Editorial rules for every table and article
Every figure links to the page it came from, and each table shows the date we retrieved it.
Link + retrieval date
When a company reports numbers about its own product, we say so, and we keep them apart from independent results.
Labeled: vendor-reported
We compare speeds only when model, quantization and backend match. Otherwise we list the rows side by side and state the differences.
Model · quantization · backend
A score without a price, or a price without a score, tells you little. We show cost per task next to the index score.
Index score + USD per task
If a source leaves out a license, a date or a configuration, we write that it was not stated instead of guessing.
Not stated means not stated
Subscribe
One email a week: the news worth reading and the numbers that moved.