Observed local runtime baseline
This is an initial environment observation, not an experiment result. The runner must repeat discovery and store the returned metadata in every recorded run.
Observed on 2026-07-17 through LM Studio at http://127.0.0.1:1234:
| Property | Observed value |
|---|---|
| Canonical study model | Qwen3.6-35B-A3B |
| OpenAI-compatible inference key | qwen/qwen3.6-35b-a3b |
| Selected variant | qwen/qwen3.6-35b-a3b@4bit |
| Runtime format | MLX |
| Quantization | 4-bit |
| Architecture reported by LM Studio | qwen3_5_moe |
| Loaded context length | 262,144 tokens |
| Loaded parallelism | 4 |
| Maximum context reported | 262,144 tokens |
| Reasoning capability/default | off or on / on |
| Tool-use capability | true |
| Model size reported | 20,429,364,306 bytes |
| LM Studio CLI commit | efce996 |
The same inventory exposed three embedding models. EMB001 selects the following available runtime:
| Property | Observed value |
|---|---|
| Display name | Qwen3 Embedding 0.6B |
| LM Studio key | text-embedding-qwen3-embedding-0.6b |
| Format and quantization | GGUF Q8_0 |
| Parameter label | 0.6B |
| Loaded context | 8,192 tokens |
| Maximum context | 32,768 tokens |
| Model size | 639,150,592 bytes |
| Vector dimension | 1,024 |
An endpoint probe embedded two different code snippets. It returned two different 1,024-dimensional vectors with L2 norms 0.99999996 and 0.99999997. LM Studio reported zero prompt and total tokens for these embedding calls, so embedding token counts must be measured independently rather than interpreting that API usage field as real zero cost.
Both /v1/models and /api/v1/models returned without error. A semantic
chat-completions probe returned visible MODEL_OK with finish_reason=stop.
That request used 20 prompt tokens and 152 completion tokens, of which 147 were
reported as reasoning tokens. This overhead is evidence that reasoning tokens
must be logged separately rather than treated as free or ignored.
The current study config pins the inference key, variant, format, quantization, loaded context length, and reasoning default. The exact LM Studio application version and host/hardware inventory still need to be captured before the pilot is frozen.