On this page
- At a glance
- What each tool is
- Supported models and how models are obtained
- Hardware and memory requirements
- Platforms and installation
- APIs and tool integrations
- Performance claims, as published by each project
- Licence and cost
- Maturity and release cadence
- Which should you choose?
- Limitations of this comparison
- Related iThub listings
- Sources
Strata and Ollama both run large language models on your own computer and serve them through a local API, but they are built for different jobs. Strata is a specialised engine for one model family: it runs Qwen3.8-Flash-Next and the smaller Qwen3.6-35B-A3B by spreading a mixture-of-experts model across the graphics card, system RAM and the SSD, so a model made for datacenter GPUs fits on a consumer PC. Ollama is a general-purpose runner: it pulls models from its library, imports GGUF and Safetensors files, runs on many GPUs or on the CPU, and offers optional cloud models. If you want Qwen3.8-Flash-Next on a desktop with a gaming GPU and 48 to 64 GB of RAM, Strata is designed for that. If you want many models, modest hardware or a large ecosystem of apps, Ollama is the more general tool.
This comparison uses only official sources: the Strata repository on GitHub (README, guides and release notes) and Ollama's website, documentation and GitHub repository, all read on 17 November 2026. iThub did not run or benchmark either tool, and every speed figure below is a project's own claim, labelled as such. For directory context, see Strata on iThub and Ollama on iThub.
At a glance
| Strata | Ollama | |
|---|---|---|
| What it is | Engine and server for Qwen mixture-of-experts models, with a desktop app and a Terminal | General runner for open models: app, CLI, REST API, optional cloud |
| Models | Qwen3.8-Flash-Next (several packs and variants) and Qwen3.6-35B-A3B | Models from the ollama.com library, plus your own GGUF or Safetensors imports |
| How models arrive | Chosen in the app or in strata launch, downloaded from Hugging Face |
ollama run or ollama pull from the library, ollama create from a Modelfile |
| Memory approach | Experts in RAM, the most used cached on the GPU, a lookup table on the SSD | On the GPU when the model fits, otherwise across GPUs or partly on the CPU |
| GPUs | NVIDIA GeForce RTX 20 to 50, AMD Radeon RX 7700 to 7900, RX 9060 and 9070, AI PRO R9700 (RX 6800 and 6900 experimental), Intel Arc on Linux, Apple Silicon, Strix Halo | NVIDIA compute capability 5.0 and newer, AMD through ROCm v7, Apple Metal, Vulkan on Windows and Linux |
| Systems | Windows 10/11 and Linux x86-64, macOS 15+ on Apple Silicon, Linux ARM64 on DGX Spark | Windows 10 22H2+, macOS 14+ (x86 Macs CPU only), Linux x86-64 and ARM64 |
| Local API | http://127.0.0.1:8080: OpenAI Chat Completions and Responses, Anthropic Messages |
http://localhost:11434: native REST API plus subsets of the OpenAI and Anthropic APIs |
| Coding tools | strata connect writes settings for OpenCode, Codex CLI, Claude Code and others |
ollama launch starts Claude Code, Codex, OpenCode and others |
| Cloud | No account; the app can add outside services such as OpenAI or OpenRouter with your key | Optional cloud models with an Ollama account; a local-only mode is available |
| Licence | MIT; each model keeps its own licence | MIT |
| Latest release | 0.2.15, 16 November 2026 | v0.40.16, 9 November 2026 |
What each tool is
Strata: one model family, spread across the whole PC
Strata is published as Niko1221/Strata on GitHub, and the repository describes it as "Qwen3.8-Flash-Next on any consumer hardware". Its How it works guide explains the idea. Qwen3.8-Flash-Next is made of 24,576 small experts, and each generated word needs only 10 of them. The graphics card caches the most requested experts, RAM holds all of them, the processor computes the ones the card lacks in parallel with the GPU, and a large lookup table stays on the SSD. The project adds that more VRAM matters more than a faster GPU, because every extra gigabyte holds about 700 more experts.
Strata ships as a desktop app and a Terminal app (strata launch). The README calls it "Free, open source, and yours to run" and says no account is needed. The engine was built with parts of llama.cpp and ggml, and since 0.2.14 it reads prompts with its own code on NVIDIA and AMD cards instead of code taken from llama.cpp. For a deeper look at the engine and its packs, read the iThub guide to Strata on consumer hardware.
Ollama: a general runner for open models, locally or in the cloud
Ollama's GitHub repository describes it as a way to "Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models." Typing ollama offers to run a model or connect agents such as Claude Code, OpenCode and Codex. The README names llama.cpp as its supported backend, and since v0.40.0 Ollama runs model architectures supported by the MLX runtime on MLX by default on Apple Silicon.
Ollama also runs models in its own cloud, which requires an Ollama account. Local use does not depend on it: the FAQ states that Ollama runs locally and does not see your prompts or data when you run locally.
Supported models and how models are obtained
Strata: two Qwen models, several packs
Strata's model guide covers Qwen3.8-Flash-Next with its variants and Qwen3.6-35B-A3B. For Qwen3.8-Flash-Next the README recommends ISTA-DASLab packs first, then Unsloth and other providers, and the model guide says this order is a recommendation, not a statement about which provider wins on quality or speed. The four ISTA-DASLab sizes are Q2_0, IQ2_XS, IQ3_XXS and IQ3_S. Variants include a Coder pack that keeps 256 of each layer's 512 experts, Swift 1.5 (a fine-tune that thinks for a shorter time), Unsloth's 4 to 8 bit files, agentionai's smaller and slower Gyro packs, and three NVFP4 versions for NVIDIA cards only.
Qwen3.6-35B-A3B is the smaller choice, offered as Unsloth GGUFs in three sizes. It runs on NVIDIA and AMD cards, while the Intel Arc engine and the Mac engine do not run it.
Model files come from Hugging Face when you install a model, according to Strata's safety and privacy page, and ModelScope can be chosen as a second source. You pick a model in the app or in strata launch, and strata models import shows what a folder or .gguf file you already have would become. Since 0.2.15, Strata loads models only on its own engine.
The model list also shrinks. Strata's DeepSeek V4.1 note says "Strata no longer runs DeepSeek V4.1-Flash" and that version 0.2.13 was the last one that set it up and started it. Files you already downloaded stay on disk, and Strata names their folder and size so you can delete them.
Ollama: a model library plus your own imports
Ollama pulls models from its library on ollama.com, which the README points to for the full list. You can also build your own: a Modelfile is described as "the blueprint to create and share customized models using Ollama", and the import guide covers Safetensors weights and GGUF files. Ollama does not quantize GGUF models during import, so quantize them beforehand with a tool such as llama.cpp's llama-quantize. Models you create can be pushed to ollama.com for others to try, and cloud models run from the same app, CLI, local API and libraries.
Where the two overlap
Both tools can run Qwen3.8-Flash-Next, packaged differently. Ollama's library page for qwen3.8-flash-next describes "a multimodal mixture-of-experts model with 125B total parameters and only 6B active per token". The model page lists a 125b-mlx tag at 105 GB, and the tags page lists a q4_K_M tag at 120 GB plus larger q8_0 and bf16 tags. Ollama's v0.40.16 release notes add that Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon.
Strata takes the other route: its model guide says Q2_0 and IQ2_XS fit on a PC with 48 GB of RAM and every ISTA-DASLab size fits with 64 GB. By contrast, Unsloth's 4 to 8 bit files of the same model "keep the experts at more bits than the sizes above, so they need more RAM". The numbers are not like for like (Ollama lists tag sizes, Strata lists the RAM a PC needs), but they show the design difference. Ollama's library also lists qwen3.6 with a 35b tag.
Hardware and memory requirements
Strata
Strata does not publish one minimum specification. Its install guide says memory and disk needs depend on the model, quantization, context and enabled features. For Qwen3.8-Flash-Next, RAM is the gating resource: the model fits when RAM is at least the experts plus about 10 GB for the system and other programs. A low-RAM mode helps machines with a big card, and the guide states that "A 32 GB PC with a 24 GB GPU runs Q2_0, IQ2_XS and the Coder this way". Setup can also read experts from the SSD, which works and is much slower. The download is 66 to 76 GB for the three smaller sizes. Qwen3.6-35B-A3B needs 32 to 52 GB of RAM.
The README lists the GPUs shown in the table above, plus NVIDIA DGX Spark on Linux. CUDA 13 packages need NVIDIA driver 580 or newer, and older NVIDIA cards use the CUDA 12 engine. A machine without a usable card is not supported: the troubleshooting guide says that when there is "no graphics card Strata can use (only an Intel UHD or Iris chip, a virtual machine, or Windows on ARM), Strata says so". On a Mac, a machine under 16 GB of unified memory is told the model does not fit. A layer split can put one model on several cards of the same make, but one engine drives one make of card, so a model is never split across NVIDIA and AMD.
Ollama
Ollama's hardware support page states that it supports NVIDIA GPUs with compute capability 5.0 and newer on driver 550 and newer. AMD cards run through ROCm v7 on Linux and Windows, Apple GPUs through the Metal API, and further GPUs on Windows and Linux through Vulkan, which is enabled by default when the backend is installed. Without a supported GPU, Ollama still runs: Intel Macs get CPU-only support, and the GPU page explains how to force CPU use.
Memory planning in Ollama follows the model you pick; the Windows guide warns that models "can be tens to hundreds of GB in size". The default context depends on VRAM: 4k below 24 GiB, 32k from 24 to 48 GiB and 256k from 48 GiB, and the context length guide advises against offloading the model to the CPU for performance. When a model does not fit on one GPU, the FAQ says it is spread across all available GPUs, and several models can stay loaded at the same time if memory allows.
Platforms and installation
Strata on Windows vs Linux
Strata has one installer per system, not one per graphics card: a Windows setup file, a Linux AppImage and a Mac disk image. On Windows the installer needs no administrator rights; on Linux the AppImage is made executable, with a fallback flag when FUSE is unavailable. On both, the app downloads the engine for your graphics card on first start, while the Mac app carries its Metal engine inside. The Terminal installs with one shell or PowerShell command, then strata launch walks you through model, context, storage and access.
Most differences between Windows and Linux concern hardware coverage. Intel Arc runs on Linux only, through the Terminal, and the DGX Spark uses the Terminal on Linux ARM64, where there is no desktop app. Strata's model and hardware table names the Radeon AI PRO R9700 in its AMD on Linux row but not in the Windows row. The Docker guide notes that Docker Desktop on Windows passes NVIDIA cards through WSL 2 but not Radeon or Arc cards. Several known problems in the 0.2.15 notes are specific to Radeon cards on Linux.
Ollama
Ollama needs Windows 10 22H2 or newer and installs in your home directory without administrator rights; a standalone zip lets you run ollama serve as a service. On macOS it needs Sonoma (version 14) or newer. On Linux a one-line script installs it, with manual packages for x86-64 and ARM64 and an extra ROCm package for AMD GPUs. The official ollama/ollama image is on Docker Hub. Its AMD coverage is narrower on Windows than on Linux: the Windows list names Radeon RX 7000 and Radeon PRO W7000 cards, and the docs recommend Vulkan as the fallback for some RX 6000 systems.
APIs and tool integrations
Strata
Strata's server listens on http://127.0.0.1:8080. The engine reference lists OpenAI Chat Completions, Anthropic Messages and the OpenAI Responses API, each with streaming and tools, plus model list, health, status and metrics endpoints. The metrics can be read as Prometheus data with vLLM's metric names, so dashboards written for vLLM can read them. strata connect covers OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent, OpenClaw and the DeepSeek Harness: it writes each tool's settings, and --check sends one prompt and one tool call the way that tool does. Access is chosen at launch: this computer only, your local network with an API key, a Cloudflare quick tunnel or Tailscale.
Ollama
Ollama's own REST API runs on port 11434, with /api/chat and /api/generate, and official Python and JavaScript libraries wrap it. The OpenAI compatibility page says Ollama supports a subset of the OpenAI API, and the Anthropic compatibility page does the same for the Anthropic Messages API, aimed at tools such as Claude Code. Local clients must send an API key value, but Ollama ignores it. ollama launch claude starts Claude Code against a chosen model, the README lists launch integrations for Codex, Copilot CLI, DeepSeek Harness, Droid and OpenCode, and a long community list covers chat interfaces, editors, SDKs and RAG tools.
Network exposure and privacy
Both bind to the local machine by default. Strata's safety page states that "There is no telemetry" and that the server listens on your PC only, with an API key needed for access from another device. Once a model is downloaded, Strata runs without internet. Ollama binds 127.0.0.1 port 11434 by default, and OLLAMA_HOST changes that. Because Ollama also offers cloud models, the FAQ documents a local-only mode: setting OLLAMA_NO_CLOUD=1 disables cloud models and web search.
Performance claims, as published by each project
Neither project's figures were checked by iThub, and the two sets cannot be compared with each other.
Strata, as stated by the project. The How it works guide says the model's built-in draft layer lets you "get the same quality answer, 1.6-1.8x sooner", and that a long document is read "at over 1,000 tokens per second". The model guide's older measurements, taken on an RTX 5070 (12 GB) with a Ryzen 5 7600 and 64 GB of RAM, report that a 32K prompt takes about 15 seconds with Q2_0. The 0.2.14 release notes report 7,870 tokens per second for a 32,000-token prompt on an RTX 5090 with UD-Q4_K_XL. Strata's documentation index adds its own caveat: "A published result on one GPU does not establish support or speed on every GPU." Community reports and a draft benchmark protocol (strata bench run) are in the repository.
Ollama, as stated by the project. The Ollama documentation and release notes read for this article do not publish tokens-per-second tables. They describe changes instead, such as the 27 GiB memory reduction for Qwen3.8 Flash Next on Apple silicon and the move to MLX in v0.40.0. ollama ps shows how much of a loaded model runs on the GPU and how much on the CPU, a useful first check when a model is slow.
Licence and cost
Strata is released under the MIT licence. Its How it works guide notes that a few bundled parts carry their own licences, that the models are not part of the repository, and that each model's own licence applies to its files. Ollama is also MIT-licensed. Neither licence covers the model weights you download, so check the licence of each model before commercial use.
Local use costs nothing in either case. Strata is free and needs no account. Ollama's pricing page lists a $0 pay-as-you-go plan that includes running models locally and charges per token for cloud models, with paid plans from $20 a month that add cloud usage credits.
Maturity and release cadence
Ollama has the longer track record. Its blog goes back to 2023, including the announcement of the official Docker image in October 2023, and its version numbers have reached v0.40.16, published on 9 November 2026 after v0.40.15 on 3 November.
Strata is young and moves quickly. Its first release, v0.1.0 on 24 September 2026, was a ready-made engine for Windows with NVIDIA RTX 30, 40 and 50 cards. By 0.2.15 on 16 November 2026, one day after 0.2.14, it also covered AMD, Intel Arc, Apple Silicon and the DGX Spark. The release notes list measured runs and known problems; the 0.2.15 notes warn, for example, that the first launch of a fresh install takes minutes while the model and its drafter are prepared. The removal of DeepSeek V4.1-Flash after 0.2.13 shows that the model list can change between releases.
Which should you choose?
You want Qwen3.8-Flash-Next on a gaming PC
Strata is built for this case: a graphics card with 12 to 24 GB and 48 to 64 GB of RAM, with the experts split across GPU, RAM and SSD. Ollama also lists the model, but its smallest listed tag is 105 GB, while Strata's two smallest ISTA-DASLab packs fit a PC with 48 GB of RAM.
You want many models, or to switch models often
Choose Ollama. Its library covers many model families, and several models can stay loaded at once if memory allows. Strata runs two Qwen models.
Your machine has 16 to 32 GB of memory, or no supported GPU
Ollama fits better. Its library includes smaller models, such as the 27B qwen3.8, listed at 18 GB, and it can run on the CPU. Strata has lighter routes, such as the Coder pack on a 32 GB PC with a 12 to 16 GB card or the Gyro-S pack at 27.6 GB of VRAM and RAM combined, at lower quality, but both still need a supported graphics card.
You connect coding agents such as Claude Code or Codex CLI
Both work: strata connect --check tests a tool against Strata's server, and ollama launch starts supported tools directly. The deciding factor is which model you want behind the agent.
You run Intel Arc on Windows, or an older GPU
Ollama's Vulkan support covers additional GPUs on Windows and Linux, and its NVIDIA support starts at compute capability 5.0. Strata runs on Intel Arc only under Linux, treats the Radeon RX 6800 and 6900 as experimental, and lists older GPUs in its guides under experimental hardware.
Neither fits
If you want a different trade-off, the iThub guide to Ollama alternatives lists other local runners, and the overview of running LLMs locally with Ollama, Strata, Xolotl, GPT4All and AnythingLLM places both tools among desktop alternatives.
Limitations of this comparison
- It reflects documentation, library pages and release notes as of 17 November 2026. Both projects publish often, Strata especially.
- iThub did not run either tool, so this article makes no claim about measured speed, answer quality or stability.
- Speed figures come from each project, on named hardware and engine versions, and are not comparable across projects.
- Memory figures differ in kind: Strata lists the RAM a PC needs for a pack, while Ollama lists the size of each tag.
- The quality of different quantizations of the same model was not assessed.
- Ollama's cloud plans, prices and model list change outside the local software.
Related iThub listings
Strata and Ollama sit in the Local LLMs and Inference category. Both use code from the llama.cpp project, which has its own llama.cpp profile.
Sources
All pages accessed on 17 November 2026.
- Strata: README, How it works, Models, Install, Model and hardware table, Engine reference, Terminal launch, Safety and privacy, Docker, Mac, Troubleshooting, Multiple GPUs, Guides index, Community benchmarks, DeepSeek V4.1 note, LICENSE, releases v0.1.0, v0.2.14 and v0.2.15github.com
- Ollama on GitHub: README, LICENSE, releases v0.40.0, v0.40.15 and v0.40.16github.com
- Ollama documentation: Hardware support, Windows, macOS, Linux, Importing a model, Modelfile, FAQ, Context length, Cloud, OpenAI compatibility, Anthropic compatibilitydocs.ollama.com
- ollama.com: qwen3.8-flash-next and its tags, qwen3.8, qwen3.6, Pricing, Blogollama.com
Frequently asked questions
What is Strata LLM?
Strata is a free, MIT-licensed local LLM engine published as Niko1221/Strata on GitHub. It runs Qwen3.8-Flash-Next and Qwen3.6-35B-A3B on consumer PCs by keeping the model's experts in system RAM, caching the most used ones on the graphics card and reading a lookup table from the SSD. It ships as a desktop app and a Terminal app for Windows, Linux and Apple Silicon Macs, needs no account, and serves OpenAI- and Anthropic-compatible APIs on http://127.0.0.1:8080 for coding tools.
What is the main difference between Strata and Ollama?
Scope. Strata is built around one model family and is designed to make Qwen3.8-Flash-Next, a mixture-of-experts model with 125B total parameters, fit on a PC with a gaming GPU and 48 to 64 GB of RAM. Ollama is a general runner: it pulls many models from its library, imports GGUF and Safetensors files, runs on a wider range of GPUs or on the CPU, and adds optional cloud models. Both are MIT-licensed and free for local use.
Can Ollama run the same Qwen3.8-Flash-Next model as Strata?
Yes. Ollama's library lists qwen3.8-flash-next, and its v0.40.16 release notes mention a memory reduction for the model on Apple silicon. The packaging differs. As of 17 November 2026, Ollama's smallest listed tag for the model is 105 GB and its q4_K_M tag is 120 GB, while Strata's model guide says its Q2_0 and IQ2_XS packs fit a PC with 48 GB of RAM, using ISTA-DASLab files that keep the experts at fewer bits.
Does Strata work the same on Windows and Linux?
Mostly. Both systems get one installer (a Windows setup file or a Linux AppImage), download the engine for the graphics card on first start and offer the same app and Terminal. The differences are in hardware: Intel Arc runs on Linux only, the NVIDIA DGX Spark is supported on Linux ARM64 through the Terminal, the hardware table names the Radeon AI PRO R9700 only for Linux, and Docker Desktop on Windows passes NVIDIA cards to containers but not Radeon or Arc cards.
Does Strata still run DeepSeek V4.1-Flash?
No. Strata's DeepSeek V4.1 note says Strata no longer runs DeepSeek V4.1-Flash and that version 0.2.13 was the last release that set it up and started it. Model files you already downloaded stay where they are, and Strata names their folder and size so you can delete them. Current releases run Qwen3.8-Flash-Next and Qwen3.6-35B-A3B. Ollama is a separate project, and its repository description names DeepSeek among the model families it runs.
Can I use Claude Code or Codex CLI with Strata or Ollama?
Yes, with both. Strata serves OpenAI Chat Completions, OpenAI Responses and Anthropic Messages on http://127.0.0.1:8080, and strata connect writes settings for OpenCode, Codex CLI, Claude Code and other tools, with a --check option that tests one prompt and one tool call. Ollama supports subsets of the OpenAI and Anthropic APIs on port 11434, and ollama launch claude starts Claude Code against a chosen local model.
Tools in this article
Open the comparison
OllamaRun open-weight LLMs locally on macOS, Windows or Linux with a REST API, plus optional paid cloud models
StrataRun large open models like Qwen3.8-Flash-Next locally on consumer GPUs, with an OpenAI-compatible API
llama.cppOpen-source C/C++ LLM inference for CPUs and GPUs, with GGUF quantization, a CLI and an OpenAI-compatible server
Facts link to their official sources. Product details can change after publication: report an error or read the editorial methodology.




