Why Local Models Are Overrated

Open-weight models are essential for AI ecosystem competition, but the hype around running them locally on consumer hardware is delusional. Models like GLM-4.2 require 400GB+ VRAM, cost $75k+ in hardware, and consume $2k/year in electricity alone. The real value of open-weight models is cloud hosting competition, not local deployment.

The Gap Between Downloadable and Runnable

Open-weight models are not local models

Models like GLM-4.2 are incredible achievements but require 400GB of VRAM in full precision (1.5TB in BF16 format) to run properly. Even quantized versions need 200GB+, making them impossible to run on any consumer hardware. Downloading a model is not the same as running it.

Unified memory is the only viable consumer path

Only three realistic consumer options exist: MacBooks with 128GB unified memory (now $3,000 extra), DGX Spark (impractical), or Strix Halo laptops/Framework desktop. MacBook prices increased 30-50% for higher RAM tiers, making them prohibitively expensive.

GPU VRAM bottleneck kills performance

A model fitting on both a MacBook and RTX 5090 runs 3x faster on the 5090. But with Deep Seek V4 Flash, the 5090 becomes bottlenecked because it lacks sufficient VRAM, forcing the model to run on system RAM instead. This creates an impossible choice: high-compute low-VRAM or low-compute high-VRAM.

Hardware Costs Are Prohibitive

Enterprise GPU pricing is absurd

RTX 5090 costs $4,300 on resale. RTX 6000 Pro with 96GB VRAM costs $13,000 (3x more expensive for identical chip, just more VRAM). A four-GPU box capable of barely running GLM-4.2 costs $75,000. These are better viewed as resale investments than functional purchases.

Electricity costs are substantial

Running an RTX 5090 24/7 in San Francisco costs approximately $5 per day, totaling $2,000 per year in electricity alone. This is just one GPU; larger deployments multiply this cost significantly. Ownership doesn't mean free operation.

Parallelism and Workflow Scaling

Real dev work requires running multiple models simultaneously

Actual workflows fluctuate between 1 and 40 parallel agents. A developer might run two threads in an IDE, spawn sub-agents to explore different code paths, or use multiple models for different tasks (orchestrator + specialized workers). Local hardware cannot scale this way; cloud infrastructure handles it trivially.

Idle hardware is wasted money

If you buy GPUs for five parallel agents but only run three, you're paying for unused capacity. If you need six agents and only have five lanes, you're blocked waiting for a workload to complete. This creates either over-provisioning waste or under-provisioning bottlenecks.

Vision and multimodal capabilities are missing

GLM-4.2, despite being an incredible coding model, has no vision capability. It cannot accept screenshots or see what it created. State-of-the-art open-weight models lack the multimodal features needed for UI work and code review feedback loops.

The Efficiency Problem

Open-weight models burn more tokens than frontier models

GLM-4.2 at $3 per million output tokens seems 10x cheaper than Opus at $25, but Opus is far more efficient. A typical Opus run costs $8 with Deep Seek while GLM-4.2 costs $4, only 2x cheaper. GLM-4.2 generates 3x more tokens to reach the same answer, negating the per-token price advantage.

Speed doesn't matter when token count triples

Even if GLM-4.2 generates output 20% faster, it requires 3x more tokens to reach the same answer. The speed gain is completely negated by the token generation overhead. This is why 3.5 Flash fails despite being fast.

GLM-4.2 is the exception, not the rule

GLM-4.2 Max is slightly cheaper than Gemini 3.5 and slightly smarter, making it a genuinely competitive option. However, GPT-5.5 on medium or low settings is neck-and-neck with GLM-4.2 while being cheaper despite higher per-token costs, due to superior efficiency.

Mobile and Edge Devices Won't Catch Up

Budget phone performance has stalled

High-end Android phones show modest improvements over 3 years, but budget Android phones ($200) have seen zero meaningful performance gains since 2022, with slight regressions. This trend will worsen in 2026 due to price hikes and manufacturing constraints.

Phone inference destroys battery and causes overheating

Running heavy models on phones drains batteries egregiously and causes thermal throttling. While on-device inference for lightweight tasks like message summaries is valuable for privacy, it's impractical for serious workloads.

Secure compute is the real privacy solution

Apple's on-device processing for lightweight tasks and secure compute enclaves offer genuine privacy benefits without requiring local model inference. These approaches are more practical than betting on device performance improvements.

The Real Value of Open-Weight Models

Competition in cloud hosting is the actual benefit

Open-weight models enable providers to compete on price and performance. OpenRouter shows this clearly: GLM-4.2 has multiple hosting options at different price points (Wafer Fast at $1.025/M tokens and 115 TPS, Deep Infra at $0.44/M tokens and 30 TPS). Closed models like Opus have no such competition.

Cloud hosting solves all local model problems

Running open-weight models on cloud providers eliminates hardware costs, electricity costs, parallelism limits, and scaling bottlenecks. The only trade-off is privacy (data sent to servers), which secure compute will eventually mitigate.

Token efficiency matters more than per-token price

When choosing between models, focus on total cost per task, not per-token pricing. A model that costs $0.03/M tokens but uses 3x more tokens is more expensive than one costing $0.25/M tokens but using 1/3 the tokens.

The Nuanced Position

Open-weight models are essential for ecosystem health

Without competitive open-weight developments like Deep Seek R1, the industry has no incentive to improve. Open-weight models are crucial for preventing monopolistic control by Anthropic and OpenAI.

Small models have legitimate use cases

Not every task requires frontier models. Lightweight models like Gemini 3.5 on phones for message summaries or weather summaries are genuinely useful. The problem is not small models; it's pretending frontier models can run locally.

Frontier models keep getting bigger

Models like Fable are estimated at 2-10 trillion parameters, far exceeding GLM-4.2. The trend is toward larger, more capable models that will never run on consumer hardware.

Notable quotes

Just because a model is open weight does not mean you can run it at home. — Theo
Open weight models really shine. They allow for competition in the hosting space. — Theo
I just think it's delusional to pretend anyone's going to run good frontier open weight models on their own hardware. — Theo
Theo - t3․gg
28 min video
3 min read
Why Local Models Are Overrated
You just saved 25 min.
The big takeaway
Open-weight models are essential for AI ecosystem competition, but the hype around running them locally on consumer hardware is delusional. Models like GLM-4.2 require 400GB+ VRAM, cost $75k+ in hardware, and consume $2k/year in electricity alone. The real value of open-weight models is cloud hosting competition, not local deployment.
The Gap Between Downloadable and Runnable
Open-weight models are not local models
Models like GLM-4.2 are incredible achievements but require 400GB of VRAM in full precision (1.5TB in BF16 format) to run properly. Even quantized versions need 200GB+, making them impossible to run on any consumer hardware. Downloading a model is not the same as running it.
GLM-4.2 Full Precision
1500 GB
GLM-4.2 BF16
400 GB
GLM-4.2 Quantized
200 GB
RTX 5090 VRAM
32 GB
VRAM requirements vs consumer GPU capacity
Unified memory is the only viable consumer path
Only three realistic consumer options exist: MacBooks with 128GB unified memory (now $3,000 extra), DGX Spark (impractical), or Strix Halo laptops/Framework desktop. MacBook prices increased 30-50% for higher RAM tiers, making them prohibitively expensive.
1
MacBook 128GB
$3,000 premium
2
Strix Halo laptop
Limited availability
3
Framework desktop
Best non-Mac option
4
DGX Spark
Impractical for inference
Consumer options for running large models
GPU VRAM bottleneck kills performance
A model fitting on both a MacBook and RTX 5090 runs 3x faster on the 5090. But with Deep Seek V4 Flash, the 5090 becomes bottlenecked because it lacks sufficient VRAM, forcing the model to run on system RAM instead. This creates an impossible choice: high-compute low-VRAM or low-compute high-VRAM.
Model fits in both VRAM pools
5090 runs 3x faster
Model exceeds 5090 VRAM
5090 bottlenecked, MacBook faster
Performance paradox: more compute doesn't help when VRAM is insufficient
Hardware Costs Are Prohibitive
Enterprise GPU pricing is absurd
RTX 5090 costs $4,300 on resale. RTX 6000 Pro with 96GB VRAM costs $13,000 (3x more expensive for identical chip, just more VRAM). A four-GPU box capable of barely running GLM-4.2 costs $75,000. These are better viewed as resale investments than functional purchases.
RTX 5090
4300 USD
RTX 6000 Pro
13000 USD
Four-GPU box
75000 USD
AMD alternative (128GB)
12000 USD
Hardware costs for local model inference
Electricity costs are substantial
Running an RTX 5090 24/7 in San Francisco costs approximately $5 per day, totaling $2,000 per year in electricity alone. This is just one GPU; larger deployments multiply this cost significantly. Ownership doesn't mean free operation.
$2,000
Annual electricity cost for one RTX 5090 in San Francisco
At $5/day, 24/7 operation
Parallelism and Workflow Scaling
Real dev work requires running multiple models simultaneously
Actual workflows fluctuate between 1 and 40 parallel agents. A developer might run two threads in an IDE, spawn sub-agents to explore different code paths, or use multiple models for different tasks (orchestrator + specialized workers). Local hardware cannot scale this way; cloud infrastructure handles it trivially.
1-40
Parallel agents in typical dev workflows
Local hardware cannot scale beyond 1-2 concurrent inferences
Idle hardware is wasted money
If you buy GPUs for five parallel agents but only run three, you're paying for unused capacity. If you need six agents and only have five lanes, you're blocked waiting for a workload to complete. This creates either over-provisioning waste or under-provisioning bottlenecks.
Vision and multimodal capabilities are missing
GLM-4.2, despite being an incredible coding model, has no vision capability. It cannot accept screenshots or see what it created. State-of-the-art open-weight models lack the multimodal features needed for UI work and code review feedback loops.
The Efficiency Problem
Open-weight models burn more tokens than frontier models
GLM-4.2 at $3 per million output tokens seems 10x cheaper than Opus at $25, but Opus is far more efficient. A typical Opus run costs $8 with Deep Seek while GLM-4.2 costs $4, only 2x cheaper. GLM-4.2 generates 3x more tokens to reach the same answer, negating the per-token price advantage.
Opus per-token price
25 $/M tokens
GLM-4.2 per-token price
3 $/M tokens
Opus typical run cost
8 USD
GLM-4.2 typical run cost
4 USD
Per-token pricing vs real-world run costs
Speed doesn't matter when token count triples
Even if GLM-4.2 generates output 20% faster, it requires 3x more tokens to reach the same answer. The speed gain is completely negated by the token generation overhead. This is why 3.5 Flash fails despite being fast.
GLM-4.2 is the exception, not the rule
GLM-4.2 Max is slightly cheaper than Gemini 3.5 and slightly smarter, making it a genuinely competitive option. However, GPT-5.5 on medium or low settings is neck-and-neck with GLM-4.2 while being cheaper despite higher per-token costs, due to superior efficiency.
Mobile and Edge Devices Won't Catch Up
Budget phone performance has stalled
High-end Android phones show modest improvements over 3 years, but budget Android phones ($200) have seen zero meaningful performance gains since 2022, with slight regressions. This trend will worsen in 2026 due to price hikes and manufacturing constraints.
High-end Android (3-year improvement)
30 % faster
Budget Android (3-year improvement)
0 % faster
iPhone (3-year improvement)
40 % faster
Performance improvements 2022-2025
Phone inference destroys battery and causes overheating
Running heavy models on phones drains batteries egregiously and causes thermal throttling. While on-device inference for lightweight tasks like message summaries is valuable for privacy, it's impractical for serious workloads.
Secure compute is the real privacy solution
Apple's on-device processing for lightweight tasks and secure compute enclaves offer genuine privacy benefits without requiring local model inference. These approaches are more practical than betting on device performance improvements.
The Real Value of Open-Weight Models
Competition in cloud hosting is the actual benefit
Open-weight models enable providers to compete on price and performance. OpenRouter shows this clearly: GLM-4.2 has multiple hosting options at different price points (Wafer Fast at $1.025/M tokens and 115 TPS, Deep Infra at $0.44/M tokens and 30 TPS). Closed models like Opus have no such competition.
Wafer Fast (GLM-4.2)
115 TPS
Fireworks Fast (GLM-4.2)
130 TPS
Friendly (GLM-4.2)
117 TPS
Deep Infra (GLM-4.2)
30 TPS
GLM-4.2 hosting options on OpenRouter
Cloud hosting solves all local model problems
Running open-weight models on cloud providers eliminates hardware costs, electricity costs, parallelism limits, and scaling bottlenecks. The only trade-off is privacy (data sent to servers), which secure compute will eventually mitigate.
1
Gap between runnable and good - solved by cloud hosting
2
Hardware costs insane - solved by renting
3
Parallelism impossible - solved by cloud scaling
4
Electricity costs high - factored into token pricing
5
Privacy concern - mitigated by secure compute
How cloud hosting addresses local model limitations
Token efficiency matters more than per-token price
When choosing between models, focus on total cost per task, not per-token pricing. A model that costs $0.03/M tokens but uses 3x more tokens is more expensive than one costing $0.25/M tokens but using 1/3 the tokens.
The Nuanced Position
Open-weight models are essential for ecosystem health
Without competitive open-weight developments like Deep Seek R1, the industry has no incentive to improve. Open-weight models are crucial for preventing monopolistic control by Anthropic and OpenAI.
Small models have legitimate use cases
Not every task requires frontier models. Lightweight models like Gemini 3.5 on phones for message summaries or weather summaries are genuinely useful. The problem is not small models; it's pretending frontier models can run locally.
Frontier models keep getting bigger
Models like Fable are estimated at 2-10 trillion parameters, far exceeding GLM-4.2. The trend is toward larger, more capable models that will never run on consumer hardware.
2-10T
Estimated parameters in Fable
Frontier models continue growing beyond local feasibility
Worth quoting
"Just because a model is open weight does not mean you can run it at home."
— Theo, at [10:13]
"Open weight models really shine. They allow for competition in the hosting space."
— Theo, at [22:01]
"I just think it's delusional to pretend anyone's going to run good frontier open weight models on their own hardware."
— Theo, at [27:07]
Made with Glimpse by Wozart
glimpse.wozart.com/v/kq1cb6cd
Share this infographic

More like this