> ## Documentation Index
> Fetch the complete documentation index at: https://docs.heylinger.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Hardware compatibility

> Check local-model choices and the memory that must remain available on your Mac or NVIDIA PC.

Linger runs local models on an Apple silicon Mac, or on an NVIDIA GPU on Windows or Linux. Setup checks your hardware and current available memory before it proposes a model. If no model fits, you can use a [Runtime with a cloud subscription](https://docs.heylinger.com/models/choose-a-runtime.md).

The model tables describe stack CLI `0.1.0.dev38` (code dated November 18, 2026). These are simulated fresh-install plans, without new speed measurements or real-device acceptance results. Packaged desktop apps include a pinned toolkit and can contain an earlier version. Check setup output for the version in your app.

## Which model does an NVIDIA PC get?

The default plan depends on your operating system, GPU memory, available system RAM, and the selected preference.

These simulated fresh-install rows assume an idle GPU whose free VRAM equals the listed capacity. The 8–16 GiB rows have 32 GiB of RAM with 28 GiB available; the 24–48 GiB rows have 64 GiB with 60 GiB available; the 96 GiB row has 128 GiB with 124 GiB available. GPU capacity alone does not guarantee the plan fits.

| Platform and total VRAM | Model | Quant | Shared context | Slots |
| - | - | - | - | - |
| Windows, 8 GiB | Spark-X2.6-4B | `Q8_0` | 64K | 2 |
| Windows, 12 GiB | Spark-X2.6-4B | `Q8_0` | 200,000 tokens | 2 |
| Linux, 8 GiB | Qwen3.6-35B-A3B | `UD-Q3_K_M` | 64K | 1 |
| Linux, 12 GiB | Qwen3.6-35B-A3B | `UD-Q4_K_M` | 64K | 1 |
| Windows or Linux, 16 GiB | Qwen3.8-27B | `UD-IQ3_S` | 64K | 2 |
| Windows or Linux, 24 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 128K | 2 |
| Windows or Linux, 32 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 200,000 tokens | 2 |
| Windows or Linux, 48 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 200,000 tokens | 2 |
| Windows or Linux, 96 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 200,000 tokens | 2 |

The Linux 8 and 12 GiB plans use [PolyStrata](https://docs.heylinger.com/models/polystrata.md) with experts in system RAM and expert computation on the CPU. The Windows rows use native llama.cpp with CUDA. The larger Linux default rows also use llama.cpp with CUDA.

Spark is a Windows fallback when no fitting validated model is available. Its catalog status remains `candidate`. Existing Windows installation checks used a 16 GiB GPU; full product acceptance on physical 8 and 12 GiB cards is still pending. The plan includes this warning.

With the `fastest` preference, the simulated NVIDIA checks at 24 GiB and above instead select Qwen3.6-35B-A3B at `UD-IQ4_XS`. This is a planner choice, not a measured speed claim.

## Which model does an Apple silicon Mac get?

The default Metal plan depends on both total unified memory and the memory available now.

These rows describe the default preference with enough available memory and a sufficient Metal working-set limit.

| Total unified memory | Model | Quant | Shared context | Slots |
| - | - | - | - | - |
| 8 GiB | No automatic local-model plan | None | None | None |
| 16 GiB | Spark-X2.6-4B | `Q4_K_M` | 64K | 2 |
| 24 GiB | Qwen3.8-27B | `UD-IQ3_S` | 64K | 2 |
| 32 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 128K | 2 |
| 48 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 200,000 tokens | 2 |
| 64 GiB | Qwen3.8-27B | `UD-Q4_K_M` | 200,000 tokens | 2 |
| 96 GiB | Qwen3.8-Flash-Next | `GSQ-RCO-IQ2_XS` | 200,000 tokens | 2 |
| 128 GiB | Qwen3.8-Flash-Next | `GSQ-RCO-IQ3_S` | 200,000 tokens | 2 |
| 256 GiB | Qwen3.8-Flash-Next | `GSQ-RCO-IQ3_S` | 200,000 tokens | 2 |

The 24 GiB Mac plan uses `q4_0` KV cache precision. In the simulated checks, 20 GiB of available memory selects Spark-X2.6-4B at `Q4_K_M` with a 64K context instead. With 21 GiB available, Qwen3.8-27B at `UD-IQ3_S` fits the default plan.

With the `fastest` preference, a 16 GiB Mac selects Spark-X2.6-1.7B at `Q8_0` and a 24 GiB Mac selects Spark-X2.6-4B at `Q4_K_M`, both with a 64K context and 2 slots. A Mac with 32 GiB or more selects Qwen3.6-35B-A3B at `UD-IQ4_XS`. This is a planner choice, not a measured speed claim.

Spark-X2.6-4B and Spark-X2.6-1.7B have no SimplexAI score, and their catalog status remains `candidate`. Lab tests ran on a physical Apple M6 with 24 GiB, not a 16 GiB Mac; full product acceptance on a physical 16 GiB Mac is still pending. The plan includes this warning.

The tables describe model selection. They make no speed promise for a chip, including another chip with the same memory capacity.

## What do the table columns mean?

Each row describes a model, its weight format, and the context capacity shared by its concurrent requests.

* **Quant** is the quantization of the model weights. The exact name identifies the weight format.
* **Shared context** is the total context pool in tokens. `K` means 1,024 tokens, so 64K is 65,536 and 128K is 131,072. A 200,000-token pool means exactly 200,000 tokens.
* **Slots** are the requests the model can serve concurrently. Two slots share the listed context pool. A 128K pool does not give each request a separate 128K window.
* **GiB** is a memory unit of 1,073,741,824 bytes. MiB means 1,048,576 bytes. Manufacturer GB labels can use a different unit.

Longer context consumes memory for the KV cache, which stores state from previous tokens. Model weights are only part of the memory requirement.

## How much memory must remain available on a PC?

The PC planner checks free VRAM and available system RAM separately.

For llama.cpp on NVIDIA, the usable GPU budget leaves a 1 GiB margin from the GPU memory available now. The model, KV cache, and runtime buffers must fit within the resulting budget.

System RAM must cover the model's host allocations, checkpoints, and caches, plus a 4 GiB system reserve and a 1 GiB gateway reserve. The combined 5 GiB reserve is not a minimum for running the model. CPU experts in the Linux 8 and 12 GiB plans also need system RAM.

For the Windows 8 and 12 GiB automatic plans, leave at least 10 GiB of system RAM available. Simulated checks put the lower bound near 9.75 GiB with the minimum cache. The recommendation rounds that estimate up.

Spark-X2.6-4B at `Q8_0` has these estimated free-VRAM requirements, including the 1 GiB margin:

| Shared context | Slots | Estimated free VRAM required |
| - | - | - |
| 64K | 2 | 7.36 GiB |
| 128K | 2 | 8.61 GiB |
| 200,000 tokens | 2 | 9.94 GiB |

For the minimum 64K plan, leave at least 7.5 GiB of VRAM free. The 8 GiB card uses that context. The 12 GiB card's default 200,000-token plan needs more free VRAM, as shown above.

These are estimates for the selected model, context, and cache. Other settings can need more memory. The saved plan's requirements take priority.

A card with 12 GiB total and only 6 GiB free does not have a 12 GiB planning budget. Another model, game, or display workload can change the choice or prevent installation.

## How much unified memory must remain available on a Mac?

macOS and the GPU use the same physical memory, so the planner does not add RAM and GPU memory together.

The macOS probe calculates available bytes as the sum of printed `Pages free`, `Pages inactive`, and `Pages speculative` from `vm_stat`, multiplied by its page size. The printed free count excludes speculative pages, so the probe includes them separately.

Wired and compressed memory do not count as available. The probe does not add purgeable or file-backed cache counters, which can overlap pages already counted. Activity Monitor groups memory differently, so its displayed remainder need not exactly match the probe.

Before the model's allocations, the Metal budget is the smaller of these limits:

* Metal's recommended maximum working set.
* Current available memory minus 2 GiB for macOS and the gateway.

The planner then subtracts a 512 MiB display reserve. Separate checks account for model weights, host allocations, KV cache, and other caches.

Under warning memory pressure, the planner holds back an additional 25% of available memory. Under critical memory pressure, planning refuses to proceed.

These simulated lower bounds describe specific configurations at normal memory pressure:

| Total unified memory and model | Shared context and slots | Current available memory |
| - | - | - |
| 16 GiB, Spark-X2.6-4B `Q4_K_M` | 64K, 2 slots | About 9 GiB available. |
| 16 GiB, Spark-X2.6-1.7B `Q8_0` | 64K, 2 slots | About 6 GiB available. |
| 24 GiB, Qwen3.8-27B `UD-IQ3_S` | 64K, 2 slots | About 21 GiB available. |
| 32 GiB, Qwen3.8-27B `UD-Q4_K_M` | Minimum 64K, 2 slots | About 25 GiB available. |
| 32 GiB, Qwen3.8-27B `UD-Q4_K_M` | 128K, 2 slots | The balanced check fits with 29 GiB available. |

Between about 6 and 9 GiB available, a Mac falls back to Spark-X2.6-1.7B. Below about 6 GiB, there is no automatic plan. The actual saved plan takes priority over these estimates.

A 16 GiB Mac with only 3.37 GB remaining does not have enough available memory for even the Spark-X2.6-1.7B agent plan. That remainder can come from subtracting App Memory, Wired Memory, and Compressed from total memory. It is not necessarily the probe's counter.

## What does setup pick when other programs use memory?

Setup steps down as available memory shrinks: first to a smaller shared context, then to a smaller quantization, then to a smaller model. Below the last step it offers no automatic plan. It never goes under a 64K context.

These steps come from the same simulated plans as the tables above. On an NVIDIA PC they follow the free VRAM on the card, with 64 GiB of RAM. When another program already holds 1 GiB or more of the card, the plan also needs your approval to share the card.

| Free VRAM | Linux | Windows |
| - | - | - |
| 25.5 GiB or more | Qwen3.8-27B `UD-Q4_K_M`, 200,000 tokens, 2 slots | Same as Linux |
| 20.5 to under 25.5 GiB | Qwen3.8-27B `UD-Q4_K_M`, 128K, 2 slots | Same as Linux |
| 18.5 to under 20.5 GiB | Qwen3.8-27B `UD-Q4_K_M`, 64K, 2 slots | Same as Linux |
| 16.5 to under 18.5 GiB | Qwen3.8-27B `UD-IQ3_S`, 128K, 2 slots | Same as Linux |
| 14.5 to under 16.5 GiB | Qwen3.8-27B `UD-IQ3_S`, 64K, 2 slots | Same as Linux |
| 11.5 to under 14.5 GiB | Qwen3.6-35B-A3B `UD-Q4_K_M` on PolyStrata, 64K, 1 slot | Spark-X2.6-4B `Q8_0`, 200,000 tokens, 2 slots |
| 10 to under 11.5 GiB | Qwen3.6-35B-A3B `UD-Q3_K_M` on PolyStrata, 64K, 1 slot | Spark-X2.6-4B `Q8_0`, 200,000 tokens, 2 slots |
| 8.5 to under 10 GiB | Qwen3.6-35B-A3B `UD-Q3_K_M` on PolyStrata, 64K, 1 slot | Spark-X2.6-4B `Q8_0`, 128K, 2 slots |
| 7.5 to under 8.5 GiB | Qwen3.6-35B-A3B `UD-Q3_K_M` on PolyStrata, 64K, 1 slot | Spark-X2.6-4B `Q8_0`, 64K, 2 slots |
| Under 7.5 GiB | No automatic plan | No automatic plan |

System RAM can also step the plan down. When the model that fits the card needs more RAM than is available, setup picks a smaller model before it refuses. These are the simulated minimums of available RAM for an idle card, with the model chosen at that minimum:

| Card | Linux | Windows |
| - | - | - |
| 8 or 12 GiB | 25 GiB, Qwen3.6-35B-A3B `UD-Q3_K_M` on PolyStrata | 9.75 GiB, Spark-X2.6-4B `Q8_0` |
| 16 GiB | 18.25 GiB, Qwen3.8-27B `UD-IQ3_S` | 9.75 GiB, Spark-X2.6-4B `Q8_0` |
| 24 GiB or more | 12 GiB, Qwen3.6-35B-A3B `UD-IQ4_XS` | 9.75 GiB, Spark-X2.6-4B `Q8_0` |

So a PC with a 16 GiB card and 16 GiB of RAM, about 12 GiB of it available, gets Spark-X2.6-4B on Windows and no automatic plan on Linux.

On a Mac, the steps follow the available unified memory. When no model with a SimplexAI score of 50 or more fits, every Mac falls back to Spark-X2.6-4B at `Q4_K_M`, then to Spark-X2.6-1.7B at `Q8_0`. Both Spark models always use a 64K context. Context sizes within a larger model can move up and down between neighboring rows, because the planner can trade KV cache precision for a larger pool.

| Available memory | Mac 16 or 24 GiB | Mac 32 GiB or more |
| - | - | - |
| 86 GiB or more | Not reachable | Qwen3.8-Flash-Next `GSQ-RCO-IQ3_S`, 64K to 200,000 tokens |
| 71.5 to under 86 GiB | Not reachable | Qwen3.8-Flash-Next `GSQ-RCO-IQ2_XS`, 64K to 200,000 tokens |
| 30.5 to under 71.5 GiB | Not reachable | Qwen3.8-27B `UD-Q4_K_M`, 200,000 tokens |
| 25 to under 30.5 GiB | Not reachable | Qwen3.8-27B `UD-Q4_K_M`, 64K to 200,000 tokens |
| 21 to under 25 GiB | Qwen3.8-27B `UD-IQ3_S`, 64K, on a 24 GiB Mac only | Qwen3.8-27B `UD-IQ3_S`, 64K to 200,000 tokens |
| 9 to under 21 GiB | Spark-X2.6-4B `Q4_K_M`, 64K | Spark-X2.6-4B `Q4_K_M`, 64K |
| 6 to under 9 GiB | Spark-X2.6-1.7B `Q8_0`, 64K | Spark-X2.6-1.7B `Q8_0`, 64K |
| Under 6 GiB | No automatic plan | No automatic plan |

All Mac rows use 2 slots. A Mac never reports more available memory than its size allows: about 13 GiB on a 16 GiB Mac, 21 GiB on 24 GiB, and 29 GiB on 32 GiB. So a 24 GiB Mac reaches Qwen3.8-27B only when almost nothing else is open, and Flash-Next needs a Mac with 96 GiB or more.

## How does setup choose and approve a plan?

Setup checks whether supported model settings fit the memory budget before it ranks the choices.

The default preference favors model quality among fitting choices, with the Spark-X2.6 fallback on a Mac when no model with a SimplexAI score of 50 or more fits. Other preferences can change the selection. The plan reports the chosen model, quantization, context, and slots.

A setup inspection labeled `basis=capacity` describes a computer's capacity after memory is freed. It does not certify that enough memory is available now. Installation still needs to pass the current-memory checks.

The planner asks for approval before model deployment. It checks memory again when you apply the plan. If the available memory has changed enough, you need a new plan. See [Local models](https://docs.heylinger.com/models/local-inference.md).

## What else does the computer need?

Local installation also needs a supported platform and enough disk space for the approved downloads.

| Item | Requirement |
| - | - |
| Mac | Apple silicon with a supported Metal runtime. Intel Macs are not supported by the desktop app. |
| PC | Windows 11 x64 or supported Linux x64, with an NVIDIA GPU for these plans. |
| Disk | Space for model files, checkpoints, and runtime downloads. The plan reports the requirements. |
| Available memory | Enough free VRAM and system RAM, or available unified memory, for the selected settings and reserves. |

If no plan fits, setup explains the refusal. It does not stop other programs. Free memory and plan again, or use your Runtime's cloud subscription.

## Keep learning

> [card] [Install the desktop app](https://docs.heylinger.com/install/desktop-app.md)
>
> Check requirements and install Linger on your computer.

> [card] [Local models](https://docs.heylinger.com/models/local-inference.md)
>
> Approve a model plan and see which engine runs it.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.