AI Compute Server Comparison for QuantEcon

[mmcky]

AI Compute Server Comparison for QuantEcon

Prepared: December 2025
Purpose: Evaluate compact AI compute options for local LLM inference to support QuantEcon translation workflows (Chinese ↔ English technical economics content)

---

Executive Summary

This report compares three leading compact AI compute platforms for running large language models locally and JAX-based computational economics research. All three options can run models like Qwen 3 (30B-70B parameters) suitable for high-quality technical translation work.

System	Price (AUD)	Memory	LLM Inference	JAX Support
Framework Desktop	~$3,200	128GB	✅ Excellent	✅ Good (ROCm)
ASUS Ascent GX10	~$7,000	128GB	✅ Excellent	✅ Best (CUDA)
Mac Studio M4 Max	~$6,000	128GB	✅ Excellent	⚠️ Limited (Metal)

Recommendation:

• For LLM translation only: The Framework Desktop offers the best value, providing equivalent memory capacity at roughly half the cost of alternatives.

• For JAX research + LLM translation: The ASUS Ascent GX10 is the best choice if JAX is critical to your workflow, despite the higher cost. Native CUDA provides the most reliable JAX experience.

• If JAX compatibility is essential but budget is constrained: The Framework Desktop with ROCm provides good JAX support at the best price point.

---

Detailed Comparison

1. Framework Desktop (AMD Ryzen AI Max+ 395 "Strix Halo")

Australian Pricing:

• 32GB configuration: ~$1,760 AUD
• 64GB configuration: ~$2,560 AUD
• 128GB configuration: ~$3,200 AUD

Hardware Specifications:

Component	Specification
Processor	AMD Ryzen AI Max+ 395 (16 Zen 5 cores, 32 threads)
GPU	Integrated Radeon 8060S (40 RDNA 3.5 CUs)
Memory	Up to 128GB LPDDR5x unified (soldered)
Memory Bandwidth	256 GB/s
Storage	2× M.2 NVMe slots (up to 16TB)
Form Factor	4.5 litres (205.5 × 96.8 × 226.1mm)
Power	Up to 140W (boost mode)
Networking	5Gb Ethernet, WiFi 7

AI/LLM Capabilities:

• Can run Qwen 3 30B-A3B (MoE) at ~34 tokens/second
• Supports quantized 70B models with 128GB configuration
• Compatible with llama.cpp, Ollama, vLLM (via ROCm)

Pros:

• Excellent price-to-memory ratio
• x86 architecture ensures broad software compatibility
• Modular design with replaceable storage, cooling, I/O ports
• Open, repairable philosophy aligns with academic values
• Strong community support

Cons:

• No CUDA support (AMD ROCm ecosystem)
• Memory is soldered (not upgradeable)
• Newer platform with less mature AI tooling than NVIDIA
• Some users report USB4/display port flakiness under Linux

Software Stack:

• llama.cpp, Ollama, vLLM (ROCm)
• Standard Linux distributions
• MLX not supported (Apple only)

---

2. ASUS Ascent GX10 (NVIDIA GB10 Grace Blackwell)

Australian Pricing:

• 1TB storage configuration: ~$7,000 AUD (ex GST)
• 4TB storage configuration: ~$8,500 AUD (estimated)

Available from DiGiCOR Australia and PB Tech

Hardware Specifications:

Component	Specification
Processor	NVIDIA GB10 Grace Blackwell Superchip
CPU	20-core ARM (10× Cortex-X925 + 10× Cortex-A725)
GPU	Blackwell GPU with 5th-gen Tensor Cores
Memory	128GB LPDDR5x unified (soldered)
Memory Bandwidth	276 GB/s
AI Performance	Up to 1 petaFLOP (FP4)
Storage	1TB / 2TB / 4TB NVMe options
Form Factor	150 × 150 × 51mm (~1.15 litres)
Power	~170W typical
Networking	10GbE + ConnectX-7 (200GbE for clustering)

AI/LLM Capabilities:

• Runs models up to 200B parameters
• Fine-tuning support for models up to 70B
• Native FP4/FP8 quantization with minimal accuracy loss
• Can cluster two units for 256GB unified memory

Pros:

• Full NVIDIA software stack (CUDA, TensorRT-LLM, vLLM)
• Best-in-class fine-tuning capabilities
• FP4 inference with near-FP8 accuracy
• Smallest form factor of all options
• Enterprise-grade networking for clustering
• NVIDIA DGX OS pre-installed

Cons:

• Highest cost option
• ARM architecture may have some software compatibility issues
• Limited storage in base configuration
• Less flexible than Framework (no modularity)

Software Stack:

• NVIDIA DGX OS (Ubuntu-based)
• CUDA, PyTorch, TensorFlow, Jupyter pre-installed
• TensorRT-LLM, vLLM optimised
• Full NVIDIA AI software stack

---

3. Apple Mac Studio (M4 Max / M3 Ultra)

Australian Pricing (apple.com/au):

M4 Max configurations:

Memory	Storage	Price (AUD)
36GB	512GB	$3,499
48GB	1TB	$4,299
64GB	1TB	$4,699
128GB	1TB	$5,899
128GB	2TB	$6,499

M3 Ultra configurations:

Memory	Storage	Price (AUD)
96GB	1TB	$6,999
256GB	2TB	$10,199
512GB	4TB	$14,599

Hardware Specifications (M4 Max 128GB):

Component	Specification
Processor	Apple M4 Max
CPU	16-core (12 performance + 4 efficiency)
GPU	40-core integrated
Neural Engine	16-core
Memory	Up to 128GB unified (M4 Max) / 512GB (M3 Ultra)
Memory Bandwidth	~546 GB/s (M4 Max) / ~819 GB/s (M3 Ultra)
Storage	Up to 8TB (M4 Max) / 16TB (M3 Ultra)
Form Factor	197 × 197mm × 95mm
Power	60-100W under AI load
Connectivity	Thunderbolt 5, 10GbE

AI/LLM Capabilities:

• M4 Max 128GB: Runs 70B models comfortably
• M3 Ultra 512GB: Can run 600B+ parameter models in memory
• ~45 tokens/second on 20B models
• Excellent MLX framework optimisation

Pros:

• Exceptional power efficiency (lowest running costs)
• Near-silent operation
• Highest memory bandwidth of consumer options
• M3 Ultra offers up to 512GB unified memory
• Excellent build quality and reliability
• Strong MLX framework for Apple Silicon

Cons:

• Apple ecosystem lock-in
• No CUDA support
• Higher cost for equivalent memory vs Framework
• Cannot fine-tune as efficiently as NVIDIA
• Limited to MLX and llama.cpp for LLM inference

Software Stack:

• MLX (Apple's optimised ML framework)
• llama.cpp (well-optimised for Apple Silicon)
• Ollama
• Standard macOS development tools

---

Small Language Model Compatibility

The following section evaluates each platform's ability to run the most capable small language models (SLMs) that can be deployed locally. These models, generally under 30B parameters, offer excellent performance on consumer hardware with modern quantization techniques.

Top Small Language Models (2025)

Based on current benchmarks and community adoption, these are the leading SLMs for local deployment:

Model	Parameters	VRAM Required (Q4)	Best For
Qwen 2.5 / Qwen 3	0.5B – 72B	0.5GB – 45GB	Multilingual, Chinese, coding
Llama 3.2 / 3.3	1B – 70B	1GB – 42GB	General purpose, instruction following
Mistral Small 3	24B	~15GB	Speed + quality balance
Phi-4	16B	~10GB	Reasoning, math, factual accuracy
Phi-4-mini	3.8B	~2.5GB	Edge/mobile, function calling
Gemma 3	4B – 27B	3GB – 17GB	Multimodal, responsible AI
DeepSeek-V3	671B (37B active)	~24GB	Coding, MoE efficiency
DeepSeek-R1	1.5B – 70B	1GB – 42GB	Reasoning, distilled models
Yi	6B – 34B	4GB – 21GB	Chinese-English bilingual
Mistral Nemo	12B	~8GB	Translation, real-time dialogue
MiniCPM	1B – 4B	1GB – 3GB	Lightweight, edge deployment

Platform Compatibility Matrix

Framework Desktop (128GB)

Model	Runnable	Quantization	Expected Speed	Notes
Qwen 3 72B	✅ Yes	Q4_K_M	~15-20 tok/s	Full model fits in memory
Qwen 3 32B	✅ Yes	Q8 / FP16	~25-35 tok/s	Excellent for translation
Qwen 3 30B-A3B (MoE)	✅ Yes	Q4_K_M	~30-35 tok/s	Recommended – best efficiency
Llama 3.3 70B	✅ Yes	Q4_K_M	~15-18 tok/s	Full model in memory
Llama 3.2 8B	✅ Yes	FP16	~80+ tok/s	Very fast
Mistral Small 3 24B	✅ Yes	Q8	~25-30 tok/s	Great speed/quality
Phi-4 16B	✅ Yes	FP16	~40-50 tok/s	Excellent reasoning
Phi-4-mini 3.8B	✅ Yes	FP16	~100+ tok/s	Real-time capable
Gemma 3 27B	✅ Yes	Q8	~25-30 tok/s	Good multimodal
DeepSeek-V3 671B	✅ Yes	Q4	~8-12 tok/s	MoE – only 37B active
DeepSeek-R1 70B	✅ Yes	Q4_K_M	~15-18 tok/s	Strong reasoning
Yi 34B	✅ Yes	Q8	~22-28 tok/s	Excellent Chinese-English
Mistral Nemo 12B	✅ Yes	FP16	~50-60 tok/s	Great for translation

Software: Ollama, llama.cpp, vLLM (ROCm)

---

ASUS Ascent GX10 (128GB)

Model	Runnable	Quantization	Expected Speed	Notes
Qwen 3 72B	✅ Yes	FP8 / FP4	~20-25 tok/s	Native FP4 support
Qwen 3 32B	✅ Yes	FP8	~35-45 tok/s	Recommended
Qwen 3 30B-A3B (MoE)	✅ Yes	FP8	~35-40 tok/s	Excellent efficiency
Llama 3.3 70B	✅ Yes	FP8 / FP4	~18-22 tok/s	Optimised via TensorRT
Llama 3.2 8B	✅ Yes	FP16	~90+ tok/s	Very fast
Mistral Small 3 24B	✅ Yes	FP8	~30-38 tok/s	Excellent speed
Phi-4 16B	✅ Yes	FP16	~50-60 tok/s	Strong reasoning
Phi-4-mini 3.8B	✅ Yes	FP16	~120+ tok/s	Real-time capable
Gemma 3 27B	✅ Yes	FP8	~28-35 tok/s	Multimodal support
DeepSeek-V3 671B	✅ Yes	FP4	~10-15 tok/s	Best MoE performance
DeepSeek-R1 70B	✅ Yes	FP8	~18-22 tok/s	Optimised reasoning
Yi 34B	✅ Yes	FP8	~28-35 tok/s	Strong bilingual
Mistral Nemo 12B	✅ Yes	FP16	~60-70 tok/s	Fast translation
Llama 3.1 405B	✅ Yes*	FP4	~5-8 tok/s	*Requires 2× GX10 cluster

Software: TensorRT-LLM, vLLM, Ollama, NVIDIA AI stack

Advantage: Native FP4/FP8 quantization via Blackwell Tensor Cores provides ~10-15% speed boost over other platforms with minimal quality loss.

---

Mac Studio M4 Max (128GB)

Model	Runnable	Quantization	Expected Speed	Notes
Qwen 3 72B	✅ Yes	Q4_K_M	~18-22 tok/s	Higher bandwidth helps
Qwen 3 32B	✅ Yes	Q8	~30-38 tok/s	Recommended
Qwen 3 30B-A3B (MoE)	✅ Yes	Q4_K_M	~32-38 tok/s	Great efficiency
Llama 3.3 70B	✅ Yes	Q4_K_M	~16-20 tok/s	Well optimised
Llama 3.2 8B	✅ Yes	FP16	~85+ tok/s	Very fast
Mistral Small 3 24B	✅ Yes	Q8	~28-34 tok/s	Good balance
Phi-4 16B	✅ Yes	FP16	~45-55 tok/s	Strong via MLX
Phi-4-mini 3.8B	✅ Yes	FP16	~110+ tok/s	Excellent edge perf
Gemma 3 27B	✅ Yes	Q8	~26-32 tok/s	Vision support via MLX
DeepSeek-V3 671B	✅ Yes	Q4	~9-13 tok/s	MoE works well
DeepSeek-R1 70B	✅ Yes	Q4_K_M	~16-20 tok/s	Good reasoning
Yi 34B	✅ Yes	Q8	~24-30 tok/s	Excellent Chinese
Mistral Nemo 12B	✅ Yes	FP16	~55-65 tok/s	Great for translation

Software: MLX, llama.cpp, Ollama

Advantage: Highest memory bandwidth (546 GB/s) improves performance on memory-bound inference. Silent operation ideal for always-on deployment.

---

Mac Studio M3 Ultra (512GB) — Extended Capability

For reference, the M3 Ultra configuration enables running the largest models:

Model	Runnable	Quantization	Expected Speed
Llama 3.1 405B	✅ Yes	Q4_K_M	~8-12 tok/s
DeepSeek-V3 671B	✅ Yes	Q8	~6-10 tok/s
Qwen 2.5 72B	✅ Yes	FP16	~25-30 tok/s

M3 Ultra starts at $6,999 AUD (96GB) and reaches $14,599 AUD (512GB)

---

Model Recommendations for QuantEcon Translation

For Chinese ↔ English technical economics translation, we recommend:

Priority	Model	Why
1st	Qwen 3 30B-A3B (MoE)	Best Chinese capability, efficient MoE architecture
2nd	Qwen 3 32B	Dense model, strong technical vocabulary
3rd	Yi 34B	Excellent Chinese-English bilingual training
4th	Mistral Nemo 12B	Fast, good translation quality
5th	DeepSeek-R1 32B	Strong reasoning for technical content

All three platforms can run the top two recommended models comfortably.

---

JAX Research Workloads

QuantEcon uses JAX extensively for computational economics research, numerical computing, and educational materials. This section evaluates each platform's suitability for JAX-based research workflows.

JAX Platform Support Summary

Platform	JAX Support	Backend	Maturity	Notes
Framework Desktop	✅ Yes	ROCm	Good	Full ROCm support since JAX 0.1.56
ASUS Ascent GX10	✅ Yes	CUDA 13	Excellent	Native CUDA, best optimised
Mac Studio	⚠️ Experimental	Metal	Limited	jax-metal plugin, not all features

---

Framework Desktop — JAX on ROCm

Support Status: ✅ Fully Supported

JAX has had full ROCm support since version 0.1.56. AMD provides official Docker images and pip packages for JAX on ROCm.

Installation:

# Docker (recommended)
docker pull rocm/jax:latest

# Or via pip (see ROCm docs for current URLs)
pip install jax[rocm]

Capabilities:

• Full JAX API support (jit, grad, vmap, pmap)
• XLA compilation via ROCm
• Mixed precision training (FP16/BF16)
• Distributed training across multiple GPUs
• Triton kernel integration

Performance Considerations:

• ROCm JAX is ~15-25% slower than CUDA JAX on equivalent hardware
• RDNA 3.5 (Strix Halo) is supported via ROCm 6.4+
• Some advanced features may lag behind CUDA versions
• AMD validates quarterly releases

QuantEcon Suitability:

Workload	Support	Performance
Numerical simulations	✅ Excellent	Near-native
Automatic differentiation	✅ Full	Good
JIT compilation	✅ Full	Good
Matrix operations	✅ Full	Good
Custom Triton kernels	✅ Yes	Requires adaptation

Recommendation: Good choice for JAX research. The 128GB unified memory allows large-scale simulations that wouldn't fit on typical GPU VRAM. Some performance overhead vs CUDA, but broad compatibility.

---

ASUS Ascent GX10 — JAX on CUDA

Support Status: ✅ Best-in-Class

The GX10 runs CUDA 13 on the Blackwell GPU, providing native, first-class JAX support. DGX OS comes with the full NVIDIA AI software stack pre-installed.

Installation:

# Pre-installed on DGX OS
pip install jax[cuda13]

Capabilities:

• Full JAX API with all optimisations
• Native XLA/CUDA compilation
• TensorRT integration for inference
• FP4/FP8 precision support via Tensor Cores
• Multi-node distributed training (2× GX10 cluster)
• NCCL for high-performance communication

Performance Considerations:

• Native CUDA = best JAX performance
• Blackwell Tensor Cores accelerate matrix operations
• 273 GB/s memory bandwidth (similar to Strix Halo's 256 GB/s)
• FP4 operations provide ~2× throughput vs FP8

QuantEcon Suitability:

Workload	Support	Performance
Numerical simulations	✅ Excellent	Best
Automatic differentiation	✅ Full	Best
JIT compilation	✅ Full	Best
Matrix operations	✅ Full	Best (Tensor Cores)
Custom Triton kernels	✅ Native	Best
Model fine-tuning	✅ Full	Best

Caveats:

• ARM CPU architecture (some x86 dependencies may need rebuilding)
• DGX OS only (Ubuntu-based, but locked to NVIDIA's distribution)
• Higher cost

Recommendation: Best platform for serious JAX research, especially if fine-tuning models or running large-scale experiments. The CUDA ecosystem maturity and Tensor Core acceleration provide the smoothest experience.

---

Mac Studio — JAX on Metal

Support Status: ⚠️ Experimental

Apple provides a Metal backend for JAX (jax-metal), but it remains experimental with significant limitations.

Installation:

pip install jax-metal

Current Status (December 2025):

• JAX Metal marked as "experimental" with warnings on every run
• Not all JAX functionality is correctly supported
• Some operations fall back to CPU
• Active development but behind CUDA/ROCm

What Works:

• Basic JAX operations (arrays, math, linear algebra)
• Simple JIT compilation
• Basic automatic differentiation
• NumPy-like API

What's Limited or Broken:

• Some advanced transformations (pmap limitations)
• Certain custom operations
• Complex control flow in JIT
• Some scientific computing libraries that depend on JAX

Better Alternative: MLX

For Apple Silicon, MLX is the preferred framework for machine learning:

• Native Apple Silicon optimisation
• NumPy/PyTorch-like API (familiar to JAX users)
• Lazy evaluation and JIT compilation
• Full unified memory utilisation
• GPU Neural Accelerator support (M5+)

However, MLX is not JAX-compatible — existing JAX code would need to be rewritten.

QuantEcon Suitability:

Workload	Support	Performance
Numerical simulations	⚠️ Partial	Variable
Automatic differentiation	⚠️ Partial	Variable
JIT compilation	⚠️ Partial	Variable
Matrix operations	✅ Good	Good
Existing JAX codebases	❌ Limited	N/A

Recommendation: Not recommended for JAX-dependent workflows. If QuantEcon's research infrastructure is built on JAX, the Mac Studio would require significant code adaptation. However, for new projects willing to adopt MLX, the Mac Studio offers excellent performance.

---

JAX Platform Comparison for QuantEcon Research

Criterion	Framework Desktop	ASUS GX10	Mac Studio
JAX Compatibility	✅ Good	✅ Excellent	⚠️ Limited
Existing JAX Code	✅ Works	✅ Works best	❌ Needs adaptation
XLA Backend	ROCm	CUDA	Metal (experimental)
Memory for Large Sims	128GB	128GB	128GB
Fine-tuning Support	⚠️ ROCm	✅ Native CUDA	❌ Limited
Distributed Training	⚠️ Possible	✅ 2-node cluster	❌ No
Ecosystem Maturity	Good	Best	Early
Price (AUD)	~$3,200	~$7,000	~$5,900

JAX Workload Recommendations

If JAX is critical to your research:

1. Best Choice: ASUS Ascent GX10 — Native CUDA provides the most reliable, performant JAX experience with full feature support.

2. Good Alternative: Framework Desktop — ROCm JAX works well for most workloads at significantly lower cost. Some performance overhead and occasional compatibility quirks.

3. Not Recommended: Mac Studio — JAX Metal is experimental and incomplete. Only consider if willing to migrate to MLX.

If flexibility is more important:

The Framework Desktop offers the best balance of JAX support, value, and x86 compatibility for running existing QuantEcon infrastructure.

---

Use Case Analysis: QuantEcon Translation Workflows

Primary Requirements

1. Run Qwen 3 models (30B-72B) for Chinese ↔ English translation
2. Handle technical economics terminology consistently
3. Process long documents (lecture series, textbooks)
4. Reliable 24/7 operation for automated translation pipelines

Model Recommendations by Platform

Platform	Recommended Model	Expected Performance
Framework Desktop	Qwen 3 30B-A3B (MoE)	~30-35 tok/s
ASUS Ascent GX10	Qwen 3 32B (FP4)	~35-45 tok/s
Mac Studio M4 Max	Qwen 3 32B (Q4)	~35-40 tok/s

All platforms can adequately handle the translation workload. Performance differences are marginal for interactive use.

Cost-Effectiveness Analysis (128GB configurations)

Platform	Price (AUD)	Cost per GB Memory	3-Year Power Cost*
Framework Desktop	~$3,200	$25/GB	~$920
ASUS Ascent GX10	~$7,000	$55/GB	~$1,120
Mac Studio M4 Max	~$5,900	$46/GB	~$630

*Estimated at $0.30/kWh, 8 hours/day operation

Fine-Tuning Considerations

If QuantEcon plans to fine-tune models on economics-specific terminology:

• Best option: ASUS Ascent GX10 (native CUDA, Unsloth support)
• Viable option: Framework Desktop (via ROCm, less optimised)
• Limited option: Mac Studio (possible but slower, less tooling)

---

Recommendations

For LLM Translation Only

Best Value: Framework Desktop (128GB) — ~$3,200 AUD

Recommended for QuantEcon if:

• Primary use is inference (translation) not fine-tuning
• Budget consciousness is important
• x86 software compatibility is valued
• JAX workloads are secondary or occasional

---

For JAX Research + LLM Translation

Best Overall: ASUS Ascent GX10 — ~$7,000 AUD

Recommended for QuantEcon if:

• JAX is critical to research workflows
• Native CUDA compatibility is essential
• Future fine-tuning on economics terminology is planned
• Clustering multiple units is anticipated
• Budget allows for premium option

Best Value with JAX: Framework Desktop (128GB) — ~$3,200 AUD

Recommended for QuantEcon if:

• JAX workloads are important but not mission-critical
• ROCm compatibility is acceptable (15-25% slower than CUDA)
• Best balance of cost and capability is desired
• x86 architecture preferred for existing infrastructure

---

Special Considerations

Mac Studio M4 Max (128GB) — ~$5,900 AUD

Consider for QuantEcon only if:

• The team already uses Apple ecosystem extensively
• Power efficiency and silent operation are priorities
• Willing to migrate JAX code to MLX (significant effort)
• JAX is not a core requirement

Not recommended if:

• Existing JAX codebases must run without modification
• CUDA/ROCm ecosystem compatibility is needed

---

Conclusion

For QuantEcon's dual requirements of LLM translation and JAX research, the platform choice depends on how critical JAX compatibility is:

If JAX is mission-critical:

The ASUS Ascent GX10 (~$7,000 AUD) is the recommended choice. Native CUDA 13 support provides the most reliable, performant JAX experience with full feature support. The higher cost is justified by:

• First-class JAX/XLA support with Tensor Core acceleration
• Full NVIDIA ecosystem for fine-tuning and distributed training
• Seamless compatibility with existing JAX codebases

If JAX is important but not mission-critical:

The Framework Desktop (~$3,200 AUD) offers the best value. ROCm JAX support works well for most workloads with ~15-25% performance overhead compared to CUDA. This option provides:

• Half the cost of NVIDIA alternatives
• Good JAX compatibility via ROCm
• Excellent LLM inference performance
• x86 architecture for broad software compatibility

If JAX is not a requirement:

The Framework Desktop remains the best value for pure LLM inference workloads. The Mac Studio is only recommended if the team is already embedded in the Apple ecosystem and willing to adopt MLX instead of JAX.

Final Recommendation for QuantEcon

Given QuantEcon's investment in JAX-based educational materials and research infrastructure, we recommend:

1. Primary choice: ASUS Ascent GX10 — if budget permits and JAX reliability is paramount
2. Budget alternative: Framework Desktop — excellent value with good JAX support via ROCm
3. Avoid: Mac Studio — JAX Metal is experimental and would require significant code migration

---

References

Hardware

• Framework Desktop: https://frame.work/au/en/desktop
• ASUS Ascent GX10: https://www.asus.com/au/networking-iot-servers/desktop-ai-supercomputer/ultra-small-ai-supercomputers/asus-ascent-gx10/
• Apple Mac Studio: https://www.apple.com/au/shop/buy-mac/mac-studio
• DiGiCOR Australia (ASUS distributor): https://www.digicor.com.au/asus-ascent-gx10

Models & Benchmarks

• Ollama Model Library: https://ollama.com/library
• Hugging Face Open Source LLMs Guide: https://huggingface.co/blog/daya-shankar/open-source-llms
• DataCamp Top Small Language Models: https://www.datacamp.com/blog/top-small-language-models
• Local LLM Hardware Guide 2025: https://www.introl.io/blog/local-llm-hardware-pricing-guide-2025
• llama.cpp Apple Silicon Performance: Performance of llama.cpp on Apple Silicon M-series ggml-org/llama.cpp#4167

JAX Documentation

• JAX Official Installation: https://docs.jax.dev/en/latest/installation.html
• JAX on ROCm: https://rocm.docs.amd.com/en/latest/compatibility/ml-compatibility/jax-compatibility.html
• ROCm JAX Installation: https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/3rd-party/jax-install.html
• JAX Metal (Apple): https://developer.apple.com/metal/jax/
• MLX Framework: https://github.com/ml-explore/mlx
• NVIDIA DGX Spark: https://www.nvidia.com/en-us/products/workstations/dgx-spark/

Prices and performance estimates current as of December 2025. Actual results may vary based on quantization, context length, and workload.

[mmcky]

The ASUS GX10 would be the best multi-purpose workload setup.