DIRECTIVE
CREATOR
Independent field guide / No. 01

We Must
Pace the
Frontier

Powerful AI. On your own machine.
A practical guide to 10 models you can download, keep and run.

Find your model Setup guide
01 / YOUR MACHINE02 / YOUR LOCAL COPY03 / A MODEL THAT FITS YOUR WORK
Qwen3.8-27B The best balance of capability and practicality.Qwen3.5-9B A smaller first download. A lower hardware bar.GLM-5.3-Flash Highest reference score in this selection.

01 / THE SHORTLIST

Find your fit.

Capability matters. So does actually being able to run it. Start with your memory budget, then open a model’s setup details.

These are starting estimates, not guaranteed fits. PC RAM and GPU VRAM are separate pools. Start at 4K–8K context and leave room for your OS.

10 OF 10 MODELSHow to read the scores ↗
01 / 10Highest reference capability

GLM-5.3-Flash

Z.ai / Unsloth Desktop

The strongest reference benchmark result in this selection. A high-end choice for difficult coding, reasoning and multi-step work.

Selected download
199.71 GB
Reference capability
42 / AA

Starting hardware
256 GB unified-memory workstation; a tight fit

The tradeoff: A roughly 200 GB download and emerging runtime support make this an enthusiast project, not a normal laptop install.

Download options Publisher & model card ↗
Specs & install for GLM-5.3-Flash
Parameters
320B total · 18B active
Max context
1,048,576 tokens
Package
UD-Q4_K_XL · community GGUF
License
MIT

Set up with Unsloth Desktop

Install Unsloth Desktop, search unsloth/GLM-5.3-Flash-GGUF, select UD-Q4_K_XL, then download and load with 8K context.

Text-chat starting route. The optional vision projector adds about 1.16 GB. Follow the model guide for its required build; generic older llama.cpp releases may not load this architecture.

Model-specific instructions ↗Install the runtime ↓
02 / 10Advanced Apple Silicon

Qwen3.8-Flash-Next

Qwen / Ollama

Strong coding and tool-use capability with a smaller download than the biggest workstation models.

Selected download
104.85 GB
Reference capability
40 / AA

Starting hardware
192 GB+ Apple Silicon; 128 GB is tight

The tradeoff: Experimental architecture and restricted commercial terms. The selected MLX package requires Apple Silicon.

Download options Publisher & model card ↗
Specs & install for Qwen3.8-Flash-Next
Parameters
125B backbone · 6B active · ~180B full package
Max context
262,144 native · up to 1M extended
Package
125b-mlx · Ollama package

Set up with Ollama

Install current Ollama on Apple Silicon. Set context to 8K in settings before running the command.

TERMINAL · AFTER INSTALL
ollama run qwen3.8-flash-next:125b-mlx

The full package includes ~51B n-gram embeddings and ~4B prediction parameters beyond its 125B backbone. The publisher license restricts commercial model-service and AI work-assistant businesses; an internal-use exception and large-scale attribution conditions apply. Read the publisher license, even if a package label says Apache.

Model-specific instructions ↗Install the runtime ↓
03 / 10Scientific reasoning alternative

DeepSeek-V4-Flash-0731

DeepSeek / llama.cpp

A strong reasoning and coding alternative, especially worth evaluating on scientific problems. The dated 0731 checkpoint is the one compared here.

Selected download
161.87 GB
Reference capability
35 / AA

Starting hardware
256 GB unified-memory workstation

The tradeoff: Long thinking sequences can be slow locally. Its 13B active count does not make the whole model small.

Download options Publisher & model card ↗
Specs & install for DeepSeek-V4-Flash-0731
Parameters
284B total · 13B active
Max context
1,048,576 tokens
Package
UD-Q8_K_XL · community GGUF
License
MIT

Set up with llama.cpp

Install a current llama.cpp build. Run this text-chat server, then open http://127.0.0.1:8080 after loading.

TERMINAL · AFTER INSTALL
llama-server -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q8_K_XL -c 8192 --jinja --temp 1.0 --top-p 1.0 --host 127.0.0.1 --port 8080

Mixed-precision source weights mean the Q8 name does not imply eight bits for every parameter. Start without the optional draft model. The reference score uses max reasoning effort; this starter command does not reproduce that evaluation.

Model-specific instructions ↗Install the runtime ↓
05 / 10Multimodal workstation option

Inkling-Small

Thinking Machines / Unsloth Desktop

A generalist with text, image and audio inputs in its native stack. Useful to explore if you already have a large workstation.

Selected download
163.27 GB
Reference capability
26 / AA

Starting hardware
256 GB unified-memory workstation

The tradeoff: “Small” is a family name, not a laptop footprint. Local applications may expose text only, and runtime support is still developing.

Download options Publisher & model card ↗
Specs & install for Inkling-Small
Parameters
276B total · 12B active (publisher figures)
Max context
1M tokens in native configuration
Package
UD-Q4_K_XL · community GGUF
License
Apache-2.0

Set up with Unsloth Desktop

Install current Unsloth Desktop. Search unsloth/Inkling-Small-GGUF, select UD-Q4_K_XL and load with 8K context. Start with text.

Use the model-specific supported app/build. The publisher reports 276B while the repository tensor counter reports about 266B; actual file totals determine this download size. Native modalities do not guarantee local wrapper support.

Model-specific instructions ↗Install the runtime ↓
06 / 10Middle workstation tier

Ling-3.0-flash

inclusionAI / llama.cpp

An efficient mixture-of-experts architecture aimed at coding, research and tool workflows, with a smaller footprint than most workstation choices.

Selected download
77.01 GB
Reference capability
25 est.

Starting hardware
128 GB unified-memory workstation

The tradeoff: The benchmark score is estimated. Promising placement, but weaker evidence than a completed independent evaluation.

Download options Publisher & model card ↗
Specs & install for Ling-3.0-flash
Parameters
124B total · 5.1B active
Max context
262,144 tokens
Package
Q4_K_M · publisher GGUF
License
MIT

Set up with llama.cpp

Install current llama.cpp with Ling architecture support. Run the server, then open http://127.0.0.1:8080 after loading.

TERMINAL · AFTER INSTALL
llama-server -hf inclusionAI/Ling-3.0-flash-GGUF:Q4_K_M -c 8192 --jinja --temp 0.6 --top-p 0.95 --host 127.0.0.1 --port 8080

Download size includes both model shards. This is the text model. Start around 8K context; the full advertised context needs substantially more working memory.

Model-specific instructions ↗Install the runtime ↓
07 / 10Personal coding experiments

MiniMax-M2.7

MiniMax / llama.cpp

A coding and tool-use model for personal self-hosted work. An option for people who want to compare a different model family.

Selected download
140.78 GB
Reference capability
23 / AA

Starting hardware
192–256 GB memory; 256 GB offers more room

The tradeoff: Substantial memory needs. Commercial use requires prior written authorization; this is not ordinary MIT licensing.

Download options Publisher & model card ↗
Specs & install for MiniMax-M2.7
Parameters
230B total · 10B active
Max context
~200K · config 204,800
Package
UD-Q4_K_XL · community GGUF

Set up with llama.cpp

For personal use, install current llama.cpp and run this text-chat server. Open http://127.0.0.1:8080 after loading.

TERMINAL · AFTER INSTALL
llama-server -hf unsloth/MiniMax-M2.7-GGUF:UD-Q4_K_XL -c 8192 --jinja --temp 1.0 --top-p 0.95 --top-k 40 --host 127.0.0.1 --port 8080

The local guide lists 196,608 context versus 204,800 in config. Start at 8K. Read the full publisher license for attribution and use conditions before adopting it for work.

Model-specific instructions ↗Install the runtime ↓
08 / 10Large multimodal generalist

MiMo-V2.5

Xiaomi / llama.cpp

Built for long-context and agent workflows, with text, image, video and audio support in its full native stack.

Selected download
191.61 GB
Reference capability
22 / AA

Starting hardware
256 GB unified-memory workstation; limited headroom

The tradeoff: A very large download for its reference score. The simple local command below is text chat, not a promise of full audio/video support.

Download options Publisher & model card ↗
Specs & install for MiMo-V2.5
Parameters
310B total · 15B active
Max context
1,048,576 tokens
Package
UD-Q4_K_XL · community GGUF
License
MIT

Set up with llama.cpp

Install current llama.cpp. Run this server, then open http://127.0.0.1:8080 after loading. Keep --jinja for the chat-template fixes.

TERMINAL · AFTER INSTALL
llama-server -hf unsloth/MiMo-V2.5-GGUF:UD-Q4_K_XL -c 8192 --jinja --temp 1.0 --top-p 0.95 --host 127.0.0.1 --port 8080

The selected conversion does not bundle a separate vision projector. Use updated runtime and quantization files; native model features and features in this local setup differ.

Model-specific instructions ↗Install the runtime ↓
09 / 10Best starting point for 16 GB

Qwen3.5-9B

Qwen / Ollama

The smallest download here. A sensible way to begin local writing, chat and lighter coding without buying a workstation.

Selected download
6.59 GB
Reference capability
14 est.

Starting hardware
16 GB unified memory, or ~12 GB VRAM + adequate RAM

The tradeoff: Lower reference capability than the larger models. Complex reasoning and long conversations are harder; its score is estimated.

Download options Publisher & model card ↗
Specs & install for Qwen3.5-9B
Parameters
9B dense model
Max context
262,144 tokens
Package
9b · Ollama Q4_K_M package
License
Apache-2.0

Set up with Ollama

Install Ollama. Begin with 4K–8K context in settings and run the command below. Keep other memory-heavy apps closed.

TERMINAL · AFTER INSTALL
ollama run qwen3.5:9b

Plan for roughly 9–12 GB for the model process at modest context, plus your OS. That is an estimate. CPU-only execution can work with enough RAM, but expect slower answers.

Model-specific instructions ↗Install the runtime ↓
10 / 10Requested specialist inclusion

Qwen3.8-27B-Uncensored

OrcaRouter / Ollama

A weight-edited Qwen derivative intended to reduce refusal behavior. Included for controlled research and comparison, not as a capability upgrade.

Selected download
17.74 GB
Reference capability
Unranked

Starting hardware
32–48 GB unified memory, or 24–32 GB VRAM

The tradeoff: No comparable independent score verified for this exact derivative. Reduced refusals do not mean better factual accuracy; read its research-use conditions.

Download options Publisher & model card ↗
Specs & install for Qwen3.8-27B-Uncensored
Parameters
27B dense derivative + vision
Max context
262,144 advertised for the model family
Package
q4_K_M · OrcaRouter Ollama package

Set up with Ollama

Read OrcaRouter’s model conditions first. For its local quantized package, install Ollama, set context to 8K, then run:

TERMINAL · AFTER INSTALL
ollama run orcarouter/Qwen3.8-27B-Uncensored:q4_K_M

Selected interpretation of “OrcaRouter Qwen uncensored”: the publisher’s guide points to this 27B family. Hugging Face downloads may require sign-in and acceptance. The hosted obsidian/Qwen3.8-27B service is separate from this Q4 local package; hosted availability and access conditions can change.

Model-specific instructions ↗Install the runtime ↓Hosted access (separate service) ↗

02 / GET IT RUNNING

Your first local model.
One step at a time.

  1. Check the memory you actually have.On a Mac: Apple menu → About This Mac. On Windows: Settings → System → About for RAM, and Task Manager → Performance → GPU for dedicated VRAM.
  2. Install a runtime.A runtime is the app that loads the model and generates replies. Ollama is the easiest starting route for the Qwen packages here.
  3. Choose one model and download it.Open “Specs & install” above. Follow the exact package and command. Leave extra disk space for the app, cache and optional components; your first download can take a while.
  4. Start small, then test your own work.Set context to 4K–8K, ask a question you can check, and watch memory. Bigger context stores more conversation but costs more memory. Download time and answer speed depend on your hardware.

Choose the right app.

Ollama: install the app and open it. Set context in settings, then paste the model’s command in Terminal or PowerShell. The first run downloads it; later runs reuse the local files. Use ollama ps to inspect loading and /bye to leave chat.

Unsloth Desktop: the documented route here for GLM and Inkling. Search the exact repository and select the named quantization. Follow each model’s guide for architecture support and operating-system requirements.

llama.cpp: an advanced local server. Install a current build for your OS/GPU. Commands assume llama-server is on your PATH; otherwise use ./llama-server on macOS/Linux or .\llama-server.exe on Windows. Run one server at a time.

Install Ollama ↗Install Unsloth ↗llama.cpp releases ↗

If it will not load, check the runtime version, exact model tag and available memory first. Reduce context and close other apps. Do not assume that a hosted model’s speed will match your computer.

Context & memory guide ↗llama.cpp build guide ↗

03 / FIELD NOTES

Know what you’re keeping.

Downloadable weights or a hosted API?

Local weights are files you keep and run on your hardware. After setup, local text generation can work offline. A hosted API runs elsewhere: it can require an account, credits and internet access, and sends requests to a provider. Hosted access is not a backup of the model. For an offline workflow, also disable cloud tools and integrations.

Does “open weights” mean “open source”?

Not automatically. Downloadable weights let you keep a copy under their license, but a fully open-source AI system has additional requirements. Licenses differ even within one model family. In particular, read the Qwen Flash-Next and MiniMax terms before commercial use. Read the Open Source Initiative definition ↗

What do B, MoE, Q4 and context mean?

B means billions of parameters — the learned numbers in a model. MoE means mixture of experts: only part of the model computes each token, but the complete weights still need storage and memory access. Q4 is a compressed weight format, not a guarantee of unchanged quality. Context is the amount of conversation and material the model can consider at once. Advertised maximums are not recommended starting settings.

Which “uncensored Qwen” is included?

This guide selects OrcaRouter’s Qwen3.8-27B-Uncensored family because its own Qwen uncensored guide identifies it. That resolves the publisher and family; a generic “uncensored Qwen” name alone does not identify an exact quantization. The selected local download is its Ollama q4_K_M package. Hosted Obsidian access is a separate service. The publisher describes research-use conditions, and the exact derivative has no independently verified comparable score here.

Can I run the maximum context or every native modality?

Not necessarily. File size is only the weights: working buffers, conversation cache, images and parallel requests need more memory. Native vision/audio/video support also depends on encoders and the runtime. The starter instructions here prioritize short text chat. CPU offloading can help a model fit, but can make it much slower.

04 / SOURCES & LIMITS

A shortlist, not a universal leaderboard.

Compared on September 13, 2026. This guide adapts the source research into exactly ten choices: eight performance picks, one smaller 16 GB option and the requested OrcaRouter derivative. It omits the older Qwen3.6-27B and the experimental G9 setup. The default order follows the Artificial Analysis Intelligence Index v4.3, with the unranked derivative last. Scores link directly to the evaluator. “Est.” means an estimated score.

Reference scores use the evaluator’s model, reasoning settings and test environment. They are not local benchmarks of these downloads. Quantization and runtime settings can change results. Small score gaps do not prove a meaningful difference for your tasks. GLM leads this selection on that reference index; Qwen3.8-27B is our editorial recommendation for everyday hardware.

Sizes are live registry-manifest or complete GGUF shard totals, rounded to decimal GB; optional projectors are excluded unless bundled by Ollama. Memory tiers are planning estimates for modest context, not measured peaks or purchase guarantees. Publisher cards, licenses, runtime guides and download records were checked. No model weights were installed or benchmarked for this guide.

Each entry links to its publisher, selected package and setup source. Community conversions are labeled. Before buying hardware, ask for a demonstration of your intended model, quantization and workload on that machine.

Download the 10-model data ↧Save an offline copy of this guide ↧New to AI coding agents? ↗