Serve LLMs with a single binary.
Serve local LLMs to your apps and coding agents through an OpenAI-compatible API. Ferrum is written in Rust and runs as a single binary, with Metal on Apple Silicon and CUDA on supported NVIDIA GPUs. The inference server needs no Python runtime.
| What would you like to do? | Start here |
|---|---|
| Run a model in your terminal | Quick Start |
| Connect an app or coding agent | Start the API server · API compatibility |
| Run Bonsai 2 27B PQ2_0 on Apple Silicon | Metal setup and tested limits |
Bonsai 2 PQ2_0 on Metal
On Apple Silicon, start a chat with Bonsai 2 27B:
Or start a local API for your apps:
The first start downloads the approximately 7.2 GB PQ2_0 weights and their
matching metadata; later starts reuse the cache. No manual file preparation is
needed. Ferrum adapts context, batching, and concurrency to available memory;
your explicit settings take precedence. The API listens at
http://127.0.0.1:8000/v1.
Tested on M1 Max / 32 GB with Metal text inference. Requires Ferrum 0.12.1 or later. Install or update Ferrum.
The shortcut selects the model files; resource configuration follows Ferrum's
normal defaults and automatic capacity handling. Planning uses the selected
device's available memory and the compiled weights, sequence states and
operation workspaces. It does not impose a
Bonsai-specific context length, batch size, concurrency, or memory budget.
Explicit configuration and CLI options take precedence. Model reasoning
behavior is preserved; append --disable-thinking if you want it off.
Ferrum keeps official Ternary Bonsai 2 27B GGUF PQ2_0
weights packed and applies the Hadamard transforms declared by the model.
Metal text inference has been validated through run and serve with FP16 KV,
including Orchestral tool execution, session continuation, and prefix-state reuse.
Testing on an M1 Max / 32 GB includes an 8K-context request. Larger contexts have
not been validated on this machine. The effective context cannot exceed the
model's declaration and may be reduced to fit the machine. Concurrent requests
share runtime capacity; the concurrency ceiling does not reserve a full context
for every request.
The alias downloads the official PQ2_0 file
and the matching config.json,
generation_config.json, tokenizer.json, tokenizer_config.json, and
chat_template.jinja from this pinned source-model revision.
CUDA PQ2_0/Hadamard operators and small mixed-state checkpoint tests have passed on a real GPU; the complete 27B CUDA model remains unvalidated. This Bonsai path does not support PTQ1_0, earlier Bonsai Q1_0/Q2_0 encodings, MLX packages, vision, or complete CPU inference. Bonsai combined with INT8 KV is not yet validated. The model's declared context limit does not establish tested coverage beyond the context above.
See it in action
Bonsai 2 27B PQ2_0 on M1 Max / 32 GB. Orchestral reads and edits Rust code through Ferrum's local API.
Watch the demo · Recording details
· Same-Mac measurements (submitted).
8× replay with every wait preserved; model already cached. The first attempt
hit Rust error E0282, followed by one repair using compiler feedback.
Both independent Rust tests then passed; the original tests and specification
were unchanged. The agent's shell execution was disabled.
Install Ferrum and Orchestral, then start the model:
In another terminal, open your project and start Orchestral:
These commands use automatic capacity settings; the recording notes include its exact configuration.
Earlier demo: Qwen3.5-9B with three concurrent agents · Recording details.
Vision
Make high-performance LLM serving simple to deploy and operate.
Quick Start
Install the latest stable Ferrum on macOS Apple Silicon or Linux x86_64:
|
The installer verifies release checksums and adds ~/.local/bin to your shell's
PATH. Open a new terminal afterward. Homebrew and manual installation
are also available.
Windows x64 supports CPU inference and compatible NVIDIA sm89 GPUs. Install from PowerShell:
irm https://ferrum.pandaailabs.com/install.ps1 | iex
The script verifies the setup checksum, installs for the current user, and adds Ferrum to PATH, including the current PowerShell session.
Installers select CPU when a supported GPU is unavailable. Package downloads use Cloudflare CDN, retain SHA256 verification, and fall back to GitHub if needed. Running the same command again installs the latest formal release.
Inspect the installed binary before downloading weights:
Run a model
With Ferrum 0.9.0 or later, use the same GGUF model on macOS, Linux, and Windows. Ferrum automatically selects the available backend; both Metal and CUDA support this Q4_K_M example.
The first run downloads about 2.55 GiB. Download time depends on your route to Hugging Face; the CLI displays download progress. Ferrum adapts context, batching, and concurrency to available memory at startup.
Serve an API
The server command is also the same on all three platforms:
Send a request from another terminal. On macOS or Linux:
In Windows PowerShell:
$body = @{ model = 'ferrum'; messages = @(@{ role = 'user'; content = 'Reply with a short hello from Ferrum.' }); max_tokens = 32 } | ConvertTo-Json -Depth 4
Invoke-RestMethod http://localhost:8000/v1/chat/completions -Method Post -ContentType 'application/json' -Body $body
Ferrum does not silently select a model. run requires MODEL, and serve
requires either --model or an intentional default_model in ferrum.toml.
A working request returns HTTP 200 with a non-empty assistant response. Native
run and serve start from the model's context limit and fit unset context,
batch and concurrency settings to currently available device memory. Startup
reports the effective limits; explicit settings are preserved or rejected with
a capacity error. The context limit includes both the rendered input and output.
The examples use --disable-thinking so the first response is short and
direct. Omit the flag to preserve the model template's default reasoning
behavior; an HTTP request can override the server default with
chat_template_kwargs.enable_thinking, Chat reasoning_effort, or Responses
reasoning.effort. See reasoning control behavior
for model support and compatibility details.
GET /v1/models also exposes optional reasoning metadata. A supported thinking
switch reports its effective default in thinking.default_enabled; explicitly
declared effort levels appear in supported_efforts. A thinking switch alone
does not imply low/medium/high levels. Omitted effort metadata means unknown support.
ferrum doctor <MODEL> resolves an alias and prints the next run and serve
commands without downloading the model or starting an inference engine.
For vNext execution, run and serve share this optional ferrum.toml setting
in the working directory:
[]
= "auto" # auto, startup, on_demand
auto uses bounded on-demand preparation on runtimes that declare support
(currently CUDA). A new shape first executes normally; later occurrences can
prepare and reuse a device program. First-use latency can therefore be higher
than steady-state latency. startup prepares the configured matrix before the
server becomes ready. Other backends retain their existing behavior; explicitly
requesting unsupported on_demand reports an error. Set reusable_execution = false
to disable device-program preparation. These options do not change request
admission, queuing, or the model's numerical profile.
KV cache precision
Ferrum v0.11.0 accepts --kv-dtype int8 in both run and serve. FP16 remains
the default. INT8 requires supported vNext standard causal attention on Metal or
portable CUDA; unsupported combinations report an error.
This reduces attention KV storage, including its quantization scales. Model
weights and fixed recurrent state retain their existing sizes. Inspect
/health → kv_storage to confirm the selected format. Whole-model checkpoint
restore requires support for every model state; resending conversation history
recomputes the input when prefix caching is disabled or no compatible checkpoint
is available. A compatible hit restores model state and processes the remaining
suffix. Session caching stores chat messages; it is separate from GPU prefix-state
reuse.
ferrum serve automatically requests prefix-state caching for the plan runtime.
The selected execution plan must support checkpointing every model state;
unsupported plans continue without caching. Retained state shares the existing
runtime memory budget and yields to foreground requests under memory pressure.
Use --disable-prefix-cache, [runtime] prefix_cache = false, or
FERRUM_PREFIX_CACHE=0 to disable it; explicit configuration and preset choices
are preserved. ferrum run leaves cross-request caching off by default.
Features
ferrum runandferrum servein one Rust binary.- OpenAI-compatible Chat Completions and stateless Responses APIs, streaming, tools, and structured output.
- Apple Silicon Metal and NVIDIA CUDA from the same runtime.
- Continuous batching, paged KV cache, prefix cache, and typed admission control.
- Optional 8-bit KV storage for supported vNext Metal and portable CUDA attention paths; FP16 remains the default.
- GGUF on Metal and CUDA; CUDA also supports GPTQ/safetensors.
- Ferrum covers language-model inference only. Supported models include Qwen3.5 4B, Qwen3.5 35B-A3B, Qwen3 30B-A3B, and Llama 3.1 8B dense.
Performance Snapshot
Latest R2 development ferrum serve checkpoint. The first three rows use
64-token input / 128-token output on Metal and 256 / 128 on CUDA. Values are
mean tok/s with the 95% confidence-interval half-width across three repeats.
| Model | M1 Max 32 GB Metal | RTX 4090 CUDA | L40S 48 GB CUDA |
|---|---|---|---|
| Qwen3.5 4B | c=16 · 61.9 ± 0.1 | c=32 · 241.3 ± 0.6 | |
| Qwen3.5 35B-A3B | c=4 · 26.1 ± 0.2 | c=16 · 174.1 ± 1.0 | |
| Qwen3 30B-A3B | c=16 · 39.6 ± 1.2 | c=32 · 214.9 ± 2.7 | |
| Qwen3.8 27B AWQ INT4 | c=4 · 78.19 ± 0.04 · c=16 · 115.12 ± 1.18 · c=32 · 115.18 ± 0.97 | ||
| Qwen3.8 27B official block-FP8 | ready 80.91 s · c=1 · 15.23 ± 0.19 · c=8 · 41.75 ± 1.26 · c=32 · 49.75 ± 0.95 | ||
| Qwen3.6 27B official block-FP8 | ready 93.39 s · c=1 · 15.15 ± 0.05 · c=8 · 42.37 ± 3.04 · c=32 · 50.38 ± 0.29 | ||
| Qwen3.6 35B-A3B official block-FP8 | ready 69.62 s · c=1 · 45.01 ± 7.54 · c=8 · 92.78 ± 2.03 · c=32 · 92.78 ± 0.84 | ||
| GPT-OSS 20B official MXFP4 | ready 23.65 s · c=1 · 61.49 ± 4.19 · c=8 · 77.16 ± 0.70 · c=32 · 77.23 ± 4.37 | ||
| Gemma 4 12B official W4A16 CT | ready 24.90 s · c=1 · 9.79 ± 0.01 · c=8 · 52.91 ± 0.88 · c=32 · 66.05 ± 6.78 |
c is active server concurrency. The first three rows completed 100 requests ×
3 repeats with zero errors.
OpenAI-Compatible API
Ferrum supports:
- chat completions and streaming usage
- stateless Responses text, reasoning replay, streaming, usage, and caller-owned function/namespace tool loops
- function tools with
auto,none,required, or a named function json_objectand strictjson_schemastructured output- multi-turn sessions, prefix cache, and session cache
- typed concurrency, memory, and scheduler controls
See OpenAI API compatibility for the exact request contract and cache product controls for prefix and session caching.
Installation
Windows 0.8.9 and later can also be installed by downloading
ferrum-<version>-windows-x86_64-cuda-sm89-setup.exe and its .sha256 file from
Releases, verifying the
checksum, and running setup. It installs under %LOCALAPPDATA%\Programs\Ferrum
and adds the current-user PATH; open a new terminal after a manual setup install.
The package includes CUDA and VC runtimes. It requires a compatible NVIDIA sm89
GPU and driver (551.78 or later); it does not install the system driver or include
models. CUDA Toolkit, Rust, and build tools are not needed. Ferrum remains a
command-line application with run and serve, without a GUI or background service.
To upgrade Windows, rerun the same PowerShell install command or the newer setup. Existing sessions keep running their original version; new launches use the updated version. Models, configuration, and existing version directories are preserved. Restart an existing server when you want it to use the update.
The macOS/Linux one-line installer selects Metal on Apple Silicon. On Linux it selects CUDA for compatible sm89 GPUs when the driver, CUDA 12.4 and NCCL runtimes can load, and otherwise selects CPU. You can require a backend or install a specific version:
|
|
To upgrade an installation made with the script, rerun the original install command. If the selected version and backend are already installed and verify successfully, the script checks the small release checksum files and skips the package download. It keeps existing version directories and switches the entry point to the verified new binary. Running sessions continue using their current version; new launches use the new version. Restart an existing server when you want it to use the update. Models and configuration are preserved.
For immediate PATH setup in the current terminal:
For Homebrew installations, use brew upgrade for the installed formula.
Homebrew 6 needs both formula definitions
trusted for its conflict check. Review them before running the trust command;
older Homebrew versions can skip it. See Homebrew's trust documentation.
# Homebrew 6: trust the reviewed formula definitions
# macOS Apple Silicon Metal
# Linux x86_64 CUDA sm89
Prebuilt tarballs from the latest stable release:
# Linux x86_64 CUDA sm89
LD_LIBRARY_PATH=/usr/local/cuda/lib64:
# macOS Apple Silicon Metal
Install the latest Metal build from crates.io:
# macOS Apple Silicon Metal
The official prebuilt Linux CUDA asset targets sm89. Linux CUDA installation requires a
compatible NVIDIA driver, CUDA runtime, and NCCL runtime on the target host.
CUDA source builds also require Ferrum's matching native-operator set, so use
the prebuilt CUDA tarball or Homebrew formula for the supported install path.
Architecture
- Contracts:
ferrum-types,ferrum-interfaces - Execution:
ferrum-engine,ferrum-scheduler,ferrum-kv,ferrum-sampler - Models and compute:
ferrum-models,ferrum-kernels,ferrum-native-ops,ferrum-quantization - Product surface:
ferrum-cli,ferrum-server,ferrum-tokenizer - Validation:
ferrum-bench-core,ferrum-testkit
Development notes: numerical execution profiles (中文).
License
MIT