Bright

Bright / Starter Shelf

Chat with a document: Apple Silicon Mac, 32 GB unified memory

An editorial starting choice with exact setup and explicit evidence limits.

Start here / Editorial starting option

Qwen3.5 4B Q4_K_M

Documentation checked 2026-10-12 · Hardware estimated, not tested

Apple Silicon Mac, 32 GB unified memory. Apple Silicon, macOS 14+, Metal.

Start at 4096 tokens, one chat. Available memory and runtime memory are not established.

Exact LM Studio searchlmstudio-community/Qwen3.5-4B-GGUF

3.383 GB selected files · Model + matching projector · Source

Apache-2.0: use under its terms. Maker and license

Download size is not runtime memory or proof of compatibility. Speed and available memory are unknown.

Open runtime documentation
Select text to copy manually

Why this starting choice: Qwen3.5 4B supplies the generative chat model for the documented LM Studio document workflow. The app handles text extraction and retrieval; an embedding-only model is not the chat choice.

Check this setup scope first.

macOS 14 or newer

Apple Silicon M series; Intel Macs excluded

Backend: metal. Starter context: 4096 tokens; one chat at a time. This is an editorial starting setting, not a measured limit.

  • This is the reviewed setup scope, not hardware detected from your browser and not a promise that the model loads or performs well.
  • Confirm installed and currently available memory separately; use one chat at a time and start with the stated context.
  • Apple unified memory is one shared pool, not separate CPU and GPU memory.
  • Record app and runtime/backend versions. Use current stable compatible runtime; historical version examples are not oldest-supported-version guarantees.

The app and engine versions are documented. Support for these exact file hashes is inferred from documented model support and the later engine, not directly tested. LM Studio 0.4.6 documents a Qwen3.5 document-chat template fix, and a maintainer confirms model support in llama.cpp engine 2.5.1. Individual loading problems were also reported. These older examples do not establish the oldest compatible version for this exact file pair.

The exact package.

Publisher: lmstudio-community. Pinned revision: f9f88ac3e234be915e23811a6d28ea287bdb927e. App and model licenses are separate.

Qwen3.5-4B-Q4_K_M.gguf / language model

2,707,513,696 transfer bytes.

SHA-256: 25082a7dd3776cc3c741c6347d3bd04523f05796607b3fbc32fa3a25dfa1418c

Pinned file evidence · Exact artifact download
mmproj-Qwen3.5-4B-BF16.gguf / projector

675,568,768 transfer bytes.

SHA-256: ae08d9d7eceb8f2d0672d61b5e6aa78b611f2942b55ed71d21414980cc454b91

Pinned file evidence · Exact artifact download

Ollama and LM Studio packages may contain different quantizations, projectors or draft models. Their byte amounts are not interchangeable.

Set it up.

  1. Check your machine conditions

    Read the separate OS, processor, backend, system RAM and dedicated VRAM conditions shown in this answer. Installed memory is not available memory. Apple Silicon uses unified memory; never add system RAM and VRAM to establish compatibility.

  2. Install and update LM Studio

    Install the current stable desktop app and update its compatible llama.cpp runtime. The source-check example is LM Studio 0.4.26 Build 4; Build 2 documents runtime 2.50.0. These are documented examples, not exact-artifact success guarantees. Use GGUF with the matching Metal or CUDA backend in your conditions; this answer does not provide an MLX path.

  3. Select the exact reviewed files

    In Discover, search the exact publisher and repository shown above. Match every selected filename and reviewed revision, including any required projector. Do not choose a similarly named curated package. If the listing cannot establish that revision, use the linked pinned files and documented import route, or stop for verification.

  4. Load one model with a short context

    Open Chat and its model loader, then select the downloaded model. Set context to 4096 tokens and use one request at a time. These are Bright editorial starting settings, not maker maxima or measured fit. My Models' settings or the pre-load configuration can set context. Keep the metadata chat template; current GGUF template settings are under Load Parameters > Advanced, but the exact current graphical labels have not been inspected.

  5. Keep Qwen3.5 reasoning settings model-specific

    Do not copy Qwen3 /think or /nothink switches. Qwen3.5 documents the enable_thinking chat-template setting. Keep the supplied metadata template and use an app-supported reasoning control only if available and understood. The exact desktop control and behavior are untested.

  6. Attach a short text-based document

    Create the fictional picnic-practice.txt below and drag it into a fresh LM Studio chat with the generative model. The documented attachment formats are .txt, .pdf and .docx. Short extracted text can enter context in full; longer documents use retrieval that can miss relevant passages. An embedding-only model is not the chat pick.

  7. Try the fictional task and verify the answer

    Use the exercise below. Compare every claimed fact or quotation with the original input. Stop if loading or extraction fails, memory pressure or swap becomes high, the system becomes unresponsive, or output repeats. Shorten input or choose a smaller reviewed option rather than forcing a fit.

  8. Keep the local-only conditions visible

    After the model files and runtime are installed, local model and document chat are documented to work offline. Search, downloads and updates use the network. Select the installed local model and avoid remote providers, LM Link, cloud services, external tools and plugins for this path. Their behavior is outside this offline claim.

  9. Estimated and untested

    Download bytes do not establish runtime memory or compatibility. Peak memory, speed and exact hardware execution remain untested. Historical application and runtime support examples are not blanket guarantees for these artifacts.

  10. Retrieval has limits

    Start with a short text-based document. Scanned PDFs may need separate OCR. Tables, figures, handwriting and full page coverage are unverified. Vision capability does not prove that the attachment extractor reads scanned pages. Verify supporting passages in the actual file; no page-count promise is made.

Try this, then check it.

In picnic-practice.txt, what happens if it rains, and who provides lunch? Quote the exact supporting sentences for each answer. If the file does not say something, say it is not stated. Create picnic-practice.txt: The Maple Street club's picnic is Saturday at noon in Riverside Park. Bring your own lunch. If it rains, meet in the community hall. RSVP to the organizer by Thursday. The club will not provide food. Check every quoted sentence against the file, then ask for the absent organizer phone number and check that the answer says it is not stated. Stop if extraction fails or supporting quotations are invented. A quotation is something to verify.

No model output is presented as a Bright test result. Stop on failed loading, sustained memory pressure, severe swapping or invented answers; verify the source and settings rather than assuming the file proves compatibility.

Other documented options to inspect

Gemma 4 12B IT Q4_K_M

Optional separately documented dense Gemma alternative on the reviewed higher-memory scope; LM Studio uses a non-draft file pair. No better-quality or speed claim.

macOS 14 or newer · 4096-token editorial starting context.

Exact checkpoint and terms

Gemma 4 26B-A4B IT QAT Q4_0

Optional separately documented QAT MoE alternative on the reviewed higher-memory scope. Active parameters do not represent total memory; no benchmark transfer or fit claim.

macOS 14 or newer · 4096-token editorial starting context.

Exact checkpoint and terms

Sources and unknowns.

  • Documentation-checked and estimated. Bright has not loaded these models or tested these computer configurations. Runtime memory, speed and output quality are unknown.
  • Download bytes are transfer size, not runtime memory. Available memory differs from installed memory. Context, app/runtime, projector, other applications and load settings affect use.
  • System RAM and dedicated VRAM are separate resources. Do not add them or infer compatibility from their sum.
  • Exact oldest compatible version of these pinned artifacts is not established. Current version mapping and historical family support are documentary evidence, not a hardware test.
  • Discover may select newer files. Compare publisher, filenames and recorded revision before treating a changed package as this reviewed option.
  • Start with 4096 tokens for this exercise. The model supports longer contexts, but longer conversations use more memory. This starting setting is not a tested memory limit or the model’s maximum context.
  • Start with a short text-based document. Scans, tables, figures and handwriting may not extract correctly; retrieval can miss passages. Check every supporting passage in the original. Offline operation excludes enabled remote tools, providers, plugins or cloud tags.

Runtime memory: not measured. Speed: not measured. No hardware test receipt exists for these documentation-led answers.

Keep this answer and its evidence.

Landscape share image · Vertical share image · This answer as JSON · Method and revision history

Local AI updates in Bright Weekly.

The existing weekly newsletter includes Bright’s editorial updates. It does not provide personalized alerts for this answer.