Local-model runtime evidence: evaluate Strata without inheriting its claims

A release is a test plan, not a result you inherit

The Strata v0.1.39 release was published on 2026-10-04 at 12:32:47 UTC. It is a dated release of a local runtime that its documentation associates with Qwen3.8-Flash-Next. That fact is useful, but it does not establish that the runtime will fit, be fast, be safe, or produce suitable answers on another machine.

The release's author measurements compare decode against 0.1.38 on an RTX 5070 using interleaved pairs. Its long-prompt case specifies VRAM and expert-cache conditions. Keep those conditions attached to the result. A measurement on one machine cannot answer the deployment question for a different machine and task. The Hacker News submitter's separate speed report is a community self-report and is not used here.

1. 1Release tag + asset digest
2. 2identify exactly what was inspected
3. 3not a safety signature

1. 1Code license
2. 2permission for runtime code
3. 3inspect separately from every weight

1. 1Model card + weight terms
2. 2decide whether a particular weight may be acquired
3. 3do not infer from code MIT

1. 1Pinned machine and workload
2. 2local repeat
3. 3record fit, latency, failures, and output checks

1. 1Any missing boundary
2. 2unknown / stop
3. 3do not substitute another runtime or cloud API

This is an original conceptual map. Its job is to prevent a common mistake: a download page, a code licence, and a throughput number answer different questions.

Three artifacts, three contracts

The repository declares an MIT licence for Strata's code. The runtime README separately says that components and every model can have their own licences. The Qwen model card's visible metadata labels the model licence as other; that is not a detailed permission grant for a particular quantization, redistribution, output, or commercial use. Read the named model card, the exact weight distribution, and their current terms before acquiring weights. MIT for the runtime code does not transfer to model weights.

v0.1.39 lists standard Windows, experimental CUDA 12, and HIP ZIPs with SHA-256 digests; some hardware paths are marked untested. Comparing a download with that value answers a byte-identity question. It does not establish trustworthy authorship, malware status, or platform suitability. For the reader's own decision, record the operating system, GPU generation, driver, RAM, VRAM, and model format before considering execution.

The release documents OpenAI-, Anthropic-, and Responses-shaped HTTP routes, including unsupported Responses features. Treat a matching route name as the beginning of a contract review. Identify the client's required fields and state behavior, compare them with the named version, and define an explicit failure when they differ. A later non-sensitive test needs its own authorized environment.

Local exposure is still an authorization boundary

The release describes a default loopback binding and documents optional broader host binding, an API key, Host/Origin checks, CORS, and opt-in MCP tools. These are author-documented controls, not a security assessment performed here. A local endpoint can still expose private prompts, tools, or files if it is deliberately bound beyond loopback, if a key is reused, or if a client is given more authority than its task needs.

The README also offers an AI-assisted setup path. That changes the trust decision: instructions can cause an agent to inspect hardware, download large weights, alter local files, start a server, or connect tools. Do not treat a repository instruction as authorization. Read the instructions first, decide the allowed filesystem, network, process, and tool scope, then perform each approved action yourself or under a bounded automation policy. This article did not download assets or weights, run a binary, start a server, call an API, or accept any licence.

Decision	Evidence needed before a real run	Stop condition
Acquire runtime code	exact tag, code licence, artifact digest, platform	licence or artifact identity is unclear
Acquire weights	exact distribution, model-card terms, disk/RAM/VRAM budget	terms or distribution revision is unclear
Start a local service	bind address, authentication, client scope, log location	service could receive non-approved data or expose tools
Compare performance	pinned runtime/model/quantization, fixed workload, repetitions, metric	hardware, prompt length, cache state, or failure denominator differs

An unexecuted N=1 fixture for your own machine

This is an offline worksheet, not an installation guide or a measured result. It uses one synthetic JSON-extraction task and does not invoke Strata, Qwen, a local server, or any API.

1. Write one record containing the intended runtime tag, candidate weight revision and licence URL, quantization file hash, operating system, GPU/driver, RAM, VRAM, context limit, prompt, expected JSON schema, and a maximum elapsed-time budget. Leave an unavailable field as unknown.
2. Prepare three synthetic inputs: one valid record, one record with a required field missing, and one record with an instruction inside the data that asks to ignore the schema. Define deterministic checks for JSON parsing, required fields, and the refusal of that embedded instruction.
3. Record three hypothetical result rows only: pass for parseable output that meets the checks; reject for malformed output or failed checks; and unknown for timeout, out-of-memory, unavailable runtime, or missing measurement. Do not average failures away.
4. If a later run is separately approved, make at most three repetitions with the same record. Retain runtime/model/quantization identifiers, warm or cold state, input and output token counts, elapsed time, peak memory when observable, output-check result, and every failure. Change one condition only after recording the baseline.

The useful conclusion is narrow: this exact configuration handled or did not handle these synthetic cases under these recorded conditions. It does not establish model quality, security, general API compatibility, cost, or permission to use real data. Connect the result to the existing quantization, KV cache, and runtime chapter for memory and throughput variables, and to the local decision-endpoint fixture for the rule that a model signal never authorizes a side effect.

There is no necessary 2024–2025 history source for this specific release-and-artifact decision. A generic local-LLM timeline would not clarify the code/weight, asset/digest, or local-binding boundaries introduced above. The dated release is the recent primary source; the undated repository, licence, and model card define current operating terms only.