8.8 KiB
Strata reference for AI tools and automated readers
Last reviewed: 25 October 2026 against the 0.2.5 release tree (app and Terminal 0.2.5; Control and engines 0.1.50). App, Control and engine versions differ; use the release notes and the installed component information. This page holds the technical context behind the short README. Use AI_SETUP.md to install Strata for a user, AGENTS.md for repository work, and MEDIA_KIT.md for articles or videos.
Product facts and sources
- Strata runs supported language models locally, with a desktop app, a terminal app, and an inference engine.
- The engine uses the GPU, system RAM and, for applicable models and memory modes, SSD storage. See how it works and engine details.
- There is one app installer per system (
Strata-0.2.9-windows-x64-setup.exe,Strata-0.2.9-linux-x86_64.AppImage,Strata-0.2.9-macos-arm64.dmg). The Windows and Linux installers carry no engine: the app downloads the engine for the graphics card on first start, or installs a file the user already has. The Mac dmg keeps its Metal engine inside. The Terminal'sstrata launchdownloads the appropriate engine and model. See installation and Terminal launch. - Strata's code is MIT licensed. Model weights and third-party components have separate licenses. See LICENSE and component/model licensing.
- Local inference does not mean every feature is offline. Downloads, updates, remote tools and user-enabled network access have separate behavior. Use SAFETY.md for privacy and network claims.
- Verify the current app and engine versions in the release notes. A screenshot, translated README or historical benchmark is not a version source.
Model families
Treat these as the three top-level families. Coder, Swift, Gyro, NVFP4 and OrcaRouter are Flash-Next variants, not additional top-level families.
| Family | Availability evidence | Catalog and limits |
|---|---|---|
| DeepSeekV4.1-Flash-1017 | deepseek41 in MULTI_FILE_FAMILIES in setup.py; Terminal since 0.2.4, desktop app since 0.2.5 |
MODELS.md: ISTA-DAS 1.75 bpw (239 GB), 2.00 bpw (257 GB) and 2.25 bpw (274 GB), 8K context, a Speed choice (Faster or Full quality). NVIDIA RTX 30, 40 and 50 (Windows, Linux), AMD Radeon AI PRO R9700 (Linux) and Ryzen AI Max / Strix Halo (Windows, gfx1151). Not on Mac, Intel Arc or Strix Halo on Linux yet. |
| Qwen3.8-Flash-Next | Catalog entries in setup.py, model guides and backend documentation | MODELS.md, QUANTS.md, GYRO.md, NVFP4.md, ORCA.md |
| Qwen3.6-35B-A3B | qwen36 in OWN_PATH_FAMILIES in setup.py |
QWEN36.md: NVIDIA and AMD; not the Intel or Mac engine |
Do not reuse Qwen quant names, model sizes or performance figures for DeepSeek. A field this page does not state is unknown, not zero and not unsupported. DeepSeek Harness is a coding tool; its presence does not establish support for a particular DeepSeek model.
Provider preference
For recommendations, documentation and agent-assisted setup, use this order:
- ISTA-DASLab: first and default recommendation when it supplies a compatible pack that fits.
- Unsloth: next choice when an appropriate ISTA-DASLab pack is unavailable or does not fit the user's needs.
- Gyro / agentionai: a memory-saving fallback; it is currently slower and must not be promoted as the fast option.
- Other providers and variants: including Swift 1.5, NVFP4 and OrcaRouter.
The maintainer works with ISTA-DASLab; this is an explicit project preference, not a measured quality ranking. Keep the user's chosen model family and explicit provider choice. Check availability, backend compatibility, memory, disk space and intended use before selecting a pack. If no ISTA-DASLab pack is documented for a family, do not invent one: Qwen3.6 has Unsloth packs, and DeepSeek has the ISTA-DAS sets.
This order is a recommendation, not the installer's default logic. A no-argument installer may preselect Swift or Gyro on some memory tiers. Read its plan and select the intended family explicitly.
Quantization and variant index
Rows follow the recommendation order. Family selectors are those documented in this checkout; confirm them
with the installed release's strata launch --help and a plan before a download.
| Parent model | Provider / variant | Family selector | Quantizations or packs |
|---|---|---|---|
| DeepSeek V4.1-Flash | ISTA-DAS | deepseek41 |
ist-das-1.75bpw (239 GB), ist-das-2.00bpw (257 GB), ist-das-2.25bpw (274 GB); Unsloth ud-q8_k_xl (508 GB, several cards) |
| Qwen3.8-Flash-Next | ISTA-DASLab, original | qwen |
Q2_0, IQ2_XS, IQ3_XXS, IQ3_S |
| Qwen3.8-Flash-Next | ISTA-DASLab, Coder | coder |
IQ1_M; pruned for coding, weaker outside code |
| Qwen3.8-Flash-Next | Unsloth | unsloth |
UD-IQ4_XS, UD-Q4_K_XL, UD-Q5_K_XL, UD-Q6_K_XL, Q8_0 |
| Qwen3.8-Flash-Next | Gyro / agentionai | gyro |
Gyro-S, Gyro-M; backend-specific validation in GYRO.md |
| Qwen3.8-Flash-Next | Swift 1.5 / UkisAI | swift |
IQ2_XS, IQ3_XXS, IQ3_S; a fine-tune, not just a quantization |
| Qwen3.8-Flash-Next | NVFP4 | nvfp4 |
NVFP4, NVFP4-NVIDIA, NVFP4-QAD; NVIDIA only |
| Qwen3.8-Flash-Next | OrcaRouter Uncensored | Manual conversion | IQ3_XXS; not an installer menu choice |
| Qwen3.6-35B-A3B | Unsloth | qwen36 |
UD-Q4_K_XL, UD-Q6_K_XL, Q8_0 |
Hardware and sizing
Read the backend guide for the exact GPU, OS, model, quantization and feature combination. A supported GPU does not imply that every model or feature works on it.
| Backend / hardware | Source |
|---|---|
| NVIDIA RTX / CUDA | INSTALL.md |
| AMD Radeon / HIP | AMD_HIP.md; RX 6000 validation differs from newer cards |
| Intel Arc / SYCL | INTEL_ARC.md; released Linux SYCL engine (Ubuntu 24.04 or newer), with model preparation and per-card limits. INTEL.md records source-port details |
| Apple Silicon / Metal | MAC.md; no DeepSeek yet |
| AMD Ryzen AI Max / Strix Halo | STRIX_HALO.md |
| Multiple GPUs and older hardware | MULTI_GPU.md, OLDER_GPUS.md |
Use the model's actual memory and disk requirements. RAM, VRAM, unified memory, download size and temporary conversion space are different quantities. Count unified memory once. Do not carry the smallest Qwen model's memory estimate over to every family. For the current planning rules, consult MODELS.md.
APIs and coding tools
Default local server address: http://127.0.0.1:8080. The configured port and access settings can differ.
| Client API | Base URL | Request endpoint |
|---|---|---|
| OpenAI Chat Completions | http://127.0.0.1:8080/v1 |
POST /v1/chat/completions |
| OpenAI Responses | http://127.0.0.1:8080/v1 |
POST /v1/responses |
| Anthropic Messages | http://127.0.0.1:8080 |
POST /v1/messages |
Use the actual configured API key when authentication is enabled. See API details. The connection registry supplies these adapters:
| Tool | strata connect identifier |
|---|---|
| Claude Code | claude-code |
| Codex CLI | codex |
| OpenCode | opencode |
| oh-my-pi | omp |
| pi.dev | pi |
| Hermes Agent | hermes |
| OpenClaw | openclaw |
| DeepSeek Harness, built into the Strata app | deepseek-harness |
strata connect lists adapters. strata connect <identifier> prints settings;
strata connect <identifier> --check --output json sends a prompt and a tool call and reports their outcomes.
Do not mistake printed configuration or an older tested version for a successful check of the current setup.
See connection instructions and limits.
Evidence and publishing
- Cite the release, commit and relevant guide behind each compatibility claim.
- Performance claims need the model, quantization, hardware, workload, engine version and source report. Keep prompt processing, decode throughput and latency separate. Do not turn a screenshot into a benchmark.
- Community benchmarks include historical reports. The standard benchmark protocol remains subject to its documented activation status; do not label legacy commands or results as implementing it.
- Mark missing or unverified information explicitly. Never invent measurements, test results or download links.
- For contribution evidence, follow CONTRIBUTING.md. For video assets and narration, follow MEDIA_KIT.md.