Keep your work close
Run inference on your own machine. Offline local workflows can keep prompts and documents on-device; connected tools have their own data flows.
MACOS · WINDOWS · LINUX · ANDROID
0.8 PREVIEWAccount…
YOUR HARDWARE. YOUR AI.Big intelligence.
Big intelligence.
Closer to home.
A gaming PC. A compact Mac. Your next AI workspace might already be on your desk. Find the model that fits—and measure what it can actually do.
TokFire Bench · Free 1–3 concurrent jobs · No login neededLOCAL INFERENCE01 / 03
LLMONE MODEL
JOB01
JOB02
JOB03
LOCALAIOUTPUT
Concurrent jobs / 3
Interactive illustration · not benchmark results
Explore real measurementsConcept hardware illustration
LOCAL MODELS.
REAL POSSIBILITIES.
REAL POSSIBILITIES.
PC Mac GGUF MLX
27BA capable model class,
within reach of personal hardware.
within reach of personal hardware.
11.8 GBPublished GSQ-RCO IQ3_S file size.
Runtime memory is additional.
Runtime memory is additional.
1 → 3Concurrent jobs to test.
Find where your setup stays useful.
Find where your setup stays useful.
WHY LOCAL. WHY NOW.Your AI should feel
Your AI should feel
like your computer.
Smaller model files open the door to useful local AI. Choose your model, keep inference close to your files, and put the hardware you own to work.
Choose how it runs
Pick your model, context and runtime. Keep a working version for your workflows and change it on your schedule.
Put your GPU to work
Use the same machine for coding, document work and chat. Local inference avoids per-token API billing, with hardware and electricity costs still yours.
THE 16 GB OPPORTUNITY27 billion parameters.
27 billion parameters.
A much smaller footprint.
ISTA-DASLab’s GSQ + RCO quantization allocates precision unevenly across a model. Its Qwen3.8-27B IQ3_S file is 11.8 GB, making a 16 GB GPU an interesting place to start.
The authors report near-baseline results on selected tests. This is compressed precision—not the original full-precision model, and not a promise of identical answers.
Read the model evidenceMODEL FILE SIZE / GBPublished file sizes · not measured VRAM usage
Then add working memory.KV cache + runtime buffers + optional vision / MTP components
THE QWEN3.8-27B SHORTLIST
One model. Six ways in.
Three PC builds. Three Macs. Curated starting configurations based on published model sizes and hardware specifications—not measured rankings.
PC · compact GGUF weights
The accessible starting point01
16 GB VRAM
RTX 5060 Ti
Modern 6–8 core CPU
32 GB RAM · 1 TB NVMe
Qwen3.8-27BGSQ-RCO IQ3_S · GGUF11.8 GB published download sizellama.cpp · CUDA
A compact 27B quant makes a 16 GB gaming GPU worth testing. Choose the 16 GB card, not the 8 GB variant.
Suggested testStart with chat, then one tool job1 job · 4K context
UFit grade pending measurement
Make more of an existing rig02
24 GB VRAM
RTX 4090
Modern 8+ core CPU
64 GB RAM · 2 TB NVMe
Qwen3.8-27BGSQ-RCO IQ3_S · GGUF11.8 GB published download sizellama.cpp · CUDA
Already own a 24 GB GPU? Keep it. The same compact model leaves more capacity for context and runtime buffers.
Suggested testTest single-job agent workflows1 job · 8K context
UFit grade pending measurement
Explore parallel workloads03
32 GB VRAM
RTX 5090
Modern 8+ core CPU
64 GB RAM · 2 TB NVMe
Qwen3.8-27BGSQ-RCO IQ3_S · GGUF11.8 GB published download sizellama.cpp · CUDA
The largest VRAM option here gives more room to explore concurrent jobs. Measure per-job latency before treating it as a shared server.
Suggested testCandidate for a 1 → 2 → 3 job sweep1 job · 8K context
UFit grade pending measurement
Mac · native MLX weights
A compact local AI desk01
48 GB unified
Mac mini · M5 Pro
18-core CPU · 20-core GPU
1 TB SSD
Qwen3.8-27BMLX community · 4-bit16.1 GB published download sizeoMLX · MLX
Unified memory gives the model and macOS a shared pool. A 48 GB configuration is a sensible starting point for this 4-bit model.
Suggested testStart with chat, then one tool job1 job · 4K context
UFit grade pending measurement
The everyday workstation02
64 GB unified
Mac Studio · M5 Max
18-core CPU · 40-core GPU
1 TB SSD
Qwen3.8-27BMLX community · 4-bit16.1 GB published download sizeoMLX · MLX
More memory for a longer working conversation and other desktop apps. Test a second job only after your single-job baseline passes.
Suggested testTest single-job agent workflows1 job · 8K context
UFit grade pending measurement
Room for a shared AI service03
96 GB unified
Mac Studio · M5 Ultra
30-core CPU · 64-core GPU
2 TB SSD
Qwen3.8-27BMLX community · 4-bit16.1 GB published download sizeoMLX · MLX
Extra capacity makes this a candidate for parallel experiments. It does not prove that every job will stay responsive.
Suggested testCandidate for a 1 → 2 → 3 job sweep1 job · 8K context
UFit grade pending measurement
Memory fit is the beginning.
Download size is not total memory use. Leave room for the OS, runtime, vision components and KV cache—the model’s working memory. Context and concurrent jobs increase demand. These starter settings are experiments, not a guarantee of speed or agent success. Use a runtime that supports this model; TokFire has not validated these six combinations yet.
Test your own setup →FROM “IT FITS” TO “IT WORKS”
Find your useful limit.
Fast text generation is only one part of the story. TokFire measures the wait, the completed work and what changes when jobs overlap.
Start with one job
Run chat, business, document or a local agent-tool workload. Capture completion, first output and task latency.
Turn up the concurrency
Repeat with 2 or 3 jobs calling the same model. Watch per-job responsiveness and failures, not just combined throughput.
Make the result useful
Read the workload-fit grade. Save the report locally, upload as a guest, and choose separately whether to publish it.
LEARN / BUILD / MEASURE
Explore the guides Make local AI part of your day.
Six practical guides: your first assistant, private document Q&A, Mac MLX, GPU memory, tool agents and concurrent jobs.
TOKFIRE BENCH / 0.9 DESKTOP PREVIEWLet your machine
Let your machine
do the talking.
Start free with 1–3 concurrent jobs on one model. Pro is planned at HK$180 once for up to 20 jobs on one activated device; paid release awaits store approval.
Methodology & limitationsmacOSApple silicon · DMGWindowsx64 · ZIP previewLinux / CLICross-platform runnerAndroidNative client · APK preview
Before you install
Install Python and a compatible inference runtime separately. Mac supports GGUF with llama.cpp or MLX with oMLX. Optional MTP requires a compatible runtime and extra configuration; TokFire does not enable it automatically.
The Mac DMG is ad-hoc signed, not notarized. Windows is unsigned; native Windows/GPU testing is pending. Android connects to an existing on-device or LAN runtime; it does not bundle a model engine. Physical-phone validation is pending.
Automatic upload is on before Run and can be disabled. Public sharing is separate and off by default. Failed uploads stay queued; reports are saved locally. Desktop 0.9 adds Pro parameter experiments with local-only reports; activation awaits product setup. Version 0.8 added CSV analysis, multi-page research and retry simulations with verified task checks. These are TokFire simulations, not actual third-party agent runs. Android 0.7 retains the original profiles.
Download sourceConnect the desktop app →