AI Compute RadarPath A — Existing PC
Use what you own.
Most people underestimate their current GPU. A 12–32 GB card runs excellent quantized models every day, free of charge.
- + 12–32 GB VRAM class
- + Ollama / LM Studio ready
- + Zero extra cost
Field guide — Issue 018Friday · 18 Sep 2026 · Models + compute decision layer
AI Right Now.
The AI ecosystem ships faster than anyone can read. We track model momentum, local hardware fit and compute prices, then turn them into one clear answer.
“What should I actually use today?”
One decision, not 50 tabs.
Today’s edition — Issue 018 · Friday · 18 Sep 2026
Since the previous edition · newest reading 2026-09-18 11:51 UTC
- Heat mover+6Phi-4 Mini · 24 → 30 · Heat Score2026-09-17 18:15 UTC → 2026-09-18 11:51 UTC
- New on the radar this week1DeepSeek-R12026-09-15 18:15 UTC → 2026-09-18 06:15 UTC
- GPU rent down−6.8%A100 · $0.88 → $0.82 per hour · Vast.ai daily median2026-09-16 23:00 UTC → 2026-09-17 23:00 UTC
- Heat mover+5Gemma 4 12B · 47 → 52 · Heat Score2026-09-17 18:15 UTC → 2026-09-18 11:51 UTC
- Cooling−4Qwen3 32B · 30 → 26 · Heat Score2026-09-17 18:15 UTC → 2026-09-18 11:51 UTC
Pick of the week — Week 37, 2026
Gemma 4 12B
Gemma 4 12B is this week's pick by rule: the largest counted Heat Score rise among tracked models that run comfortably on a consumer card, from 45 to 62 since the evening of 7 September. Its trending score is back where it stood in late August, its measured Q4_K_M file is 6.6 GB, and it runs with room to spare on any 16 GB card at 8K context.
- ▲Qwen3.8 27B · HF trending 608
- ▲Qwen3.8 Flash-Next · HF trending 226
- ▲Llama 3.1 8B · HF trending 216
- ▲GLM-5.3-Flash · HF trending 163
- ◆RTX 4090 $0.53/h
- ▲RTX 5090 $0.67/h
- ◆A100 $0.71/h
- ▲H100 $3.14/h
- ◆H200 $4.74/h
- ▲B200 $8.75/h
00 / Radar view
Every tracked model on one scope.
A contact appears at the center when its Hugging Face repository is created and drifts outward as it ages. Brightness is its Heat Score. Bearing groups models by maker and means nothing else.
How to read the scope
- Rings, from the center: 30 days, 6 months, 1 year, 2 years since the repository was created. The rim is three years and beyond.
- Brightness: Heat Score 0 to 100. A dim outline is a model still collecting its first week of history.
- Pulsing ring: repository created within the last 30 days.
- Bearing: maker sector, nothing more. Blips open the model page.
Newest contacts
- GLM-5.3Zhipu AI · repo created 2026-08-25 · 24 d agoHeat 70
- GLM-5.3-FlashZhipu AI · repo created 2026-08-25 · 24 d agoHeat 76
- Qwen3.8 Flash-NextAlibaba Qwen · repo created 2026-08-24 · 25 d agoHeat 79
- Ornith 1.5 35B-A3BOrnith AI · repo created 2026-08-18 · 31 d agoHeat 72
- Ornith 1.5 397BOrnith AI · repo created 2026-08-18 · 31 d agoHeat 42
Strongest returns
- Qwen3.8 27BAlibaba Qwen · 44 d ago83
- Llama 3.1 8BMeta · 792 d ago83
- Gemma 4 31BGoogle · 191 d ago80
Ages come from Hugging Face repository creation dates; the Heat Score from measured signals. How the Heat Score is measured
01 / Model market
What is winning
| Rank | Model | Heat0–100 · composite | TrendingHugging Face | DownloadsHugging Face | $/M inOpenRouter |
|---|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibaba Qwen · OpenGeneral · Coding · Agents · ResearchHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 608$/M in: $0.21 | 608 | 7.5M+1.8% rising | $0.21 | |
| 2 | Llama 3.1 8BMeta · OpenGeneral · CodingHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 216 | 216 | 5.9M+4.6% rising | — | |
| 3 | Gemma 4 31BGoogle · OpenGeneral · ResearchHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 50$/M in: $0.09 | 50 | 9.1M+4.2% rising | $0.09 | |
| 4 | Qwen3.8 Flash-NextAlibaba Qwen · OpenGeneral · Coding · AgentsHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 226$/M in: $0.15 | 226 | 706K+25.2% rising | $0.15 | |
| 5 | Qwen3 8BAlibaba Qwen · OpenGeneral · Coding · AgentsHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 62.5$/M in: $0.12 | 62.5 | 13.1M+1.4% rising | $0.12 | |
| 6 | Llama 3.2 3BMeta · OpenGeneralHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 29$/M in: $0.05 | 29 | 1.8M+17.4% rising | $0.05 | |
| 7 | GLM-5.3-FlashBreakoutZhipu AI · OpenGeneral · Coding · AgentsHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 163$/M in: $0.09 | 163 | 2.4M+139.1% rising | $0.09 | |
| 8 | Gemma 4 26B-A4BGoogle · OpenGeneral · ResearchHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 18$/M in: $0.09 | 18 | 9.6M+8.4% rising | $0.09 | |
| 9 | Ornith 1.5 35B-A3BOrnith AI · OpenGeneral · Coding · AgentsHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 43 | 43 | 394K+25.0% rising | — | |
| 10 | GLM-5.3CoolingZhipu AI · OpenGeneral · Agents · ResearchHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 47$/M in: $1.40 | 47 | 838K+51.9% rising | $1.40 | |
| 11 | gpt-oss 20BOpenAI · OpenGeneral · Agents · ResearchHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 26$/M in: $0.03 | 26 | 6.7M+1.3% rising | $0.03 | |
| 12 | Ornith 1.5 9BOrnith AI · OpenGeneral · CodingHugging Face ↗ (External link)GGUF ↗ (External link)Trending: 21 | 21 | 506K+35.4% rising | — |
Top 12 of 46 tracked models, ranked by Heat Score; models still collecting history follow, ranked by trending.
Heat is a composite of measured signals only: each component is the model's percentile among models with a full week of history — trending level (35%), 7-day download growth (30%), 7-day trending change (15%), 30-day downloads (15%) and Hub likes (5%). Missing components renormalize the weights and lower the shown confidence; nothing is guessed. Trending and downloads come from the Hugging Face Hub — the download counter is a rolling 30-day window, not unique users — and input pricing from OpenRouter (CC BY 4.0). The full formula, thresholds and flag rules are on the methodology page. How Heat is computed → Every tracked model has its own page →
02 / Compute ladder
Own, Spark or rent?
Models need machines. The right machine depends on how often you push past your own hardware — not on the biggest number on a spec sheet.
Path B — Personal AI box
Own the middle.
DGX Spark and GB10-class systems put 128 GB of unified memory on your desk. Rational when large local models are your daily routine.
- + 128 GB unified memory
- + 273 GB/s — capacity, not speed
- + Private by default
Path C — Elastic compute
Rent the spike.
Occasional heavy job? Rent an H100 for the afternoon instead of buying hardware that idles the rest of the year.
- + H100 / A100 / B200 class
- + Pay per hour
- + No commitment
Median $/h
- RTX 4090$0.53steady
- RTX 5090$0.67rising
- A100$0.71steady
- H100$3.14rising
- H200$4.74steady
- B200$8.75rising
03 / My setup
Your machine, your answer
◇ Estimated— Estimated: computed from our curated model and hardware catalog — not a live reading.
Local first. — Fit score 73
Local first.
- Main model
- GLM-4.7-Flash · 31.2B MoE · Q4_K_M
- Compute
- RTX 4090 · 24 GB VRAM
- Burst option
- Vast.ai when the job outgrows you
- Estimated memory
- ~18.9 GB
Why this route
- + Weights ~17.1 GB (Q4_K_M), context ~0.4 GB at 8K tokens, runtime ~1.4 GB.
- + Owning the inference beats paying per token at this frequency
Hugging Face ↗ (External link)
Estimated from quantized model file sizes plus runtime overhead and a safety margin. Real usage varies with context length.
Tap the bear to wake it
From the channel
Short answers, on YouTube
The same numbers as on this site in videos under a minute: does this model fit that card, and why. Measured file sizes, computed context memory, no speed claims.
- Rent or Buy a GPU for Local AI? ↗2026-09-17
- Gemma 4 12B: This Week's Local AI Pick ↗2026-09-15
- Your Model Fits, Until the Chat Gets Long ↗2026-09-14
04 / Hardware radar
Models need machines.
◇ Estimated— Estimated: computed from our curated model and hardware catalog — not a live reading.
Modelsneedmachines.
We track the machines that matter for local AI — consumer GPUs, personal AI systems and the rentable heavy metal.
Hardware watchlist
| Device | Memory | Class | Signal |
|---|---|---|---|
| RTX 4070Entry local | 12 GB | Steadysteady | |
| RTX 3060 12GBBudget classic | 12 GB | Huge install basesteady | |
| RTX 4080Solid local | 16 GB | Steadysteady | |
| RTX 5080Current 16 GB | 16 GB | Current genrising | |
| RTX 4060 Ti 16GBBudget 16 GB | 16 GB | Value picksteady | |
| RTX 5060 Ti 16GBCurrent budget 16 GB | 16 GB | Current genrising | |
| RTX 5070 TiFast 16 GB | 16 GB | Current genrising | |
| RTX 3090Used-market value | 24 GB | Demand risingrising | |
| RTX 4090Local workhorse | 24 GB | Demand risingrising | |
| RTX 5090Top consumer | 32 GB | Demand risingrising | |
| Mac mini M4 Pro 64GBApple unified 64 GB | 64 GB U | Apple Siliconrising | |
| Mac Studio M4 Max 64GBApple unified 64 GB | 64 GB U | Apple Siliconrising | |
| Mac Studio M3 Ultra 96GBApple unified 96 GB | 96 GB U | Apple Siliconrising | |
| NVIDIA DGX SparkPersonal AI system | 128 GB U | New categoryrising | |
| A100 80 GB (rental)Rental workhorse | 80 GB | $/h fallingfalling | |
| H100 80 GB (rental)Rental flagship | 80 GB | $/h trackedrising | |
| B200 180 GB (rental)Rental frontier | 180 GB | $/h fallingfalling |
Radar Card of the day
RTX 4090
In 2022 this machine shipped with 24 GB. Today 22 of 46 tracked models run on it at 8K context.