Caleta Private AI Appliance / Model catalogue
The model catalogue
Pick a model, get a proven size and the best regions to run it
Recommendations come only from configurations we have run on the shipping image; region rankings come from spot data we measure daily.
1
Pick a model
| Model | Download | Measured speed | Tested on | Runs from | |||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Phi-4 dense 14B | 10 GB | 95.35 t/s Q4_K_M · 1x A100 80 GB · 17 Aug 2026 16K native context is the small print on this one. | NC24ads_A100_v4NC80adis_H100_v5 2 deployments›
| $0.16/hr 1x Tesla T4 16 GB not timed there | |||||||||||||||||||||||||
gpt-oss-20b MoE reasoner, 4B active also on Arm | 13 GB | 195.47 t/s MXFP4 · 1x A100 80 GB · 17 Aug 2026 80.2 / 380 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 Also proven to load and serve on a single Tesla T4 16 GB, not timed there. | NC24ads_A100_v4NC80adis_H100_v5E32ps_v6 2 deployments›
| $0.16/hr 1x Tesla T4 16 GB not timed there | |||||||||||||||||||||||||
ERNIE-4.5-21B-A3B MoE general, 3B active also on Arm | 13 GB | 74.2 / 388 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 Arm plan. Tiny and Apache-2.0; fits the smallest Cobalt size. | E32ps_v6 | $0.16/hr 1x Tesla T4 16 GB not timed there | |||||||||||||||||||||||||
Devstral-Small-2 dense 24B | 15 GB | 59.84 t/s UD-Q4_K_XL · 1x A100 80 GB · 17 Aug 2026 It does load and serve on a single Tesla T4 16 GB, but only with the context window cut to around 6,900 tokens, so the entry price above uses the next card up rather than quietly charging you the context. | NC24ads_A100_v4NC80adis_H100_v5 2 deployments›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Qwen3-Coder-30B MoE coder, 3B active also on Arm | 18 GB | 167.89 t/s UD-Q4_K_XL · 1x A100 80 GB · 17 Aug 2026 72.8 / 833 t/s decode / prefill · UD-Q4_K_XL · E64ads_v7, no GPU, ik_llama.cpp · 14 Sep 2026 60.46 t/s UD-Q4_K_XL · 64 vCPU x86, no GPU · 23 Aug 2026 · CPU beta, shipped image, mainline llama.cpp 74.3 / 337 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 | NC24ads_A100_v4NC80adis_H100_v5E32ps_v6E64ads_v7D64ads_v7 2 deployments›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Gemma-4-26B-A4B MoE, 4B active | 18 GB | 110.38 t/s UD-Q4_K_XL · 1x A100 80 GB · 17 Aug 2026 | NC24ads_A100_v4NC80adis_H100_v5 2 deployments›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Qwen3.8-27B dense 27B | 18 GB | 47.48 t/s UD-Q4_K_XL · 1x A100 80 GB · 18 Aug 2026 Newest weights at the same speed and price as Qwen3.6-27B. No Caleta Score has been run for it, and none is carried across from Qwen3.6: same family and size class is not the same model. | NC24ads_A100_v4 1 deployment›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Qwen3.6-27B dense 27B | 18 GB | 46.65 t/s UD-Q4_K_XL · 1x A100 80 GB · 17 Aug 2026 The highest Caleta Score on the coder board. | NC24ads_A100_v4NC80adis_H100_v5 2 deployments›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
GLM-4.7-Flash MoE, 3B active | 19 GB | 115.03 t/s Q4_K_M · 1x A100 80 GB · 17 Aug 2026 | NC24ads_A100_v4NC80adis_H100_v5 2 deployments›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Qwen3-30B-A3B-Instruct MoE general, 3B active also on Arm | 19 GB | 71.3 / 337 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 Arm plan. The fast general-purpose daily driver on Cobalt. | E32ps_v6 | $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Qwen2.5-Coder-32B dense 32B | 20 GB | 42.53 t/s Q4_K_M · 1x A100 80 GB · 17 Aug 2026 | NC24ads_A100_v4NC80adis_H100_v5 2 deployments›
| $0.28/hr Quarter RTX PRO 6000 Blackwell not timed there extra setup | |||||||||||||||||||||||||
Qwen3.6-35B-A3B MoE, 3B active also on Arm | 23 GB | 136.45 t/s UD-Q4_K_XL · 1x A100 80 GB · 17 Aug 2026 54.9 / 1,045 t/s decode / prefill · UD-Q4_K_XL · E64ads_v7, no GPU, ik_llama.cpp · 14 Sep 2026 65.5 / 445 t/s Arm · decode / prefill · UD-Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 · thinking off | NC24ads_A100_v4NC80adis_H100_v5E32ps_v6E64ads_v7 2 deployments›
| $0.35/hr Azure Cobalt E32ps_v6, 251 GB RAM, no GPU | |||||||||||||||||||||||||
Qwen3-Next-80B-A3B MoE general, 3B active also on Arm | 49 GB | 59.8 / 322 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 Arm plan. Smart rather than fast. | E32ps_v6 | $0.35/hr Azure Cobalt E32ps_v6, 251 GB RAM, no GPU | |||||||||||||||||||||||||
Qwen3-Coder-Next-80B MoE coder, 3B active | 50 GB | 110.27 t/s UD-Q4_K_XL · 1x A100 80 GB · 17 Aug 2026 46.3 / 779 t/s decode / prefill · UD-Q4_K_XL · E64ads_v7, no GPU, ik_llama.cpp · 14 Sep 2026 | NC24ads_A100_v4NC80adis_H100_v5E64ads_v7 2 deployments›
| $0.78/hr 1x A100 80 GB | |||||||||||||||||||||||||
gpt-oss-120b MoE reasoner, 5B active also on Arm | 64 GB | 138 t/s MXFP4 · 1x A100 80 GB · 17 Aug 2026 · independently reproduced five weeks apart 48.7 / 768 t/s decode / prefill · MXFP4 · E64ads_v7, no GPU, ik_llama.cpp · 14 Sep 2026 54.8 / 232 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 | NC24ads_A100_v4E32ps_v6E64ads_v7 1 deployment›
| $0.35/hr Azure Cobalt E32ps_v6, 251 GB RAM, no GPU | |||||||||||||||||||||||||
GLM-4.5-Air MoE mid, 12B active also on Arm | 73 GB | 24.6 / 91 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 Arm plan. MIT licence; needs E16ps_v6 or larger. | E32ps_v6 | $0.35/hr Azure Cobalt E32ps_v6, 251 GB RAM, no GPU | |||||||||||||||||||||||||
Mistral-Small-4 MoE general, 6.5B active also on Arm | 74 GB | 113.45 / 1,476 t/s decode / prefill · UD-Q4_K_M · 1x A100 80 GB · 4 Oct 2026 · image 0.15.1; card full, context fitted by the engine to about 428K tokens 160.49 / 2,824.8 t/s decode / prefill · UD-Q4_K_M · 1x H100 NVL 94 GB · 4 Oct 2026 · image 0.15.1; card full, context fitted by the engine to about 1M tokens 30.7 / 193.6 t/s decode / prefill · UD-Q4_K_M · E64ads_v7, no GPU, llama.cpp · 4 Oct 2026 · image 0.15.1; 132 GB RAM in use at the full ~1M context 30.65 / 38.1 t/s Arm · decode / prefill · UD-Q4_K_M · Azure Cobalt E32ps_v6, llama.cpp · 3 Oct 2026 · image 0.3.3; 32K context, 70 GB RAM in use Apache-2.0. Text only on the appliance; reasoning off by default. Standard llama.cpp only: no ik_llama.cpp variant on either architecture. | E64ads_v7NC24ads_A100_v4NC40ads_H100_v5E32ps_v6 3 deployments›
| $0.35/hr Azure Cobalt E32ps_v6, 251 GB RAM, no GPU | |||||||||||||||||||||||||
Qwen3-235B MoE, 22B active also on Arm | 471 GB | 44.41 t/s BF16, unquantised · 8x MI300X · 10 Aug 2026 13.9 / 142 t/s decode / prefill · UD-Q4_K_XL · E64ads_v7, no GPU, ik_llama.cpp · 14 Sep 2026 12.8 / 49 t/s Arm · decode / prefill · Q4_K_M · Azure Cobalt E32ps_v6, ik_llama.cpp · 14 Sep 2026 The 4-bit cut (135 GB) ships on 2x H100 NVL: fit is proven, serving speed not yet published for that pairing. | NC80adis_H100_v5ND96is_MI300X_v5E32ps_v6E64ads_v7 2 deployments›
| $0.35/hr Azure Cobalt E32ps_v6, 251 GB RAM, no GPU | |||||||||||||||||||||||||
DeepSeek-V4-Flash-0731 MoE reasoner, 13B active | 155 GB | 54.3 t/s MXFP4 · 2x H100 NVL · 16 Aug 2026 12.2 / 180 t/s decode / prefill · MXFP4 · E64ads_v7, no GPU, ik_llama.cpp · 14 Sep 2026 14.06 t/s MXFP4 · 64 vCPU x86, no GPU · 23 Aug 2026 · CPU beta, shipped image, mainline llama.cpp 13.84 t/s Arm · MXFP4 · Azure Cobalt D64ps_v6 · 24 Aug 2026 155 GB does not fit one H100, so two cards is its reference box, not an upsell. | NC80adis_H100_v5D64ps_v6E64ads_v7D64ads_v7 1 deployment›
| $0.97/hr E64ads_v7, 512 GB RAM, no GPU | |||||||||||||||||||||||||
GLM-5.2 MoE frontier, 40B active also on Arm | 802 GB | 38.61 t/s Q8_0 · 8x MI300X · 10 Aug 2026 27.7 t/s UD-IQ2_M · 4x A100 80 GB · 29 Jun 2026 ~11.7 t/s Arm · UD-Q4_K_XL · Azure Cobalt E64ps_v6 · 20 Jul 2026 A 744B frontier model. The IQ3_S cut scores highest of anything we have measured; the scores were taken before the reasoning-budget change and should be re-run before being requoted. | NC96ads_A100_v4ND96is_MI300X_v5E64ps_v6 4 deployments›
| $0.97/hr E64ads_v7, 512 GB RAM, no GPU not timed there | |||||||||||||||||||||||||
DeepSeek-R1-0528 MoE reasoner, 37B active | 252 GB | ~9.1 t/s UD-Q2_K_XL · 64 vCPU x86, no GPU · 11 Jul 2026 · CPU beta The Q8 cut on 8x MI300X is loaded and coherence-checked, but its decode has no clean timing run yet. Two-bit quality is unmeasured, so no score is quoted. | E64ads_v7 2 deployments›
| $0.97/hr E64ads_v7, 512 GB RAM, no GPU | |||||||||||||||||||||||||
Qwen3-Coder-480B MoE coder, 35B active | 511 GB | 51.4 t/s Q8_0 · 8x MI300X · 10 Aug 2026 | ND96is_MI300X_v5 1 deployment›
| $12.02/hr 8x MI300X 1.5 TB | |||||||||||||||||||||||||
Kimi-K2.7-Code MoE coder, 1.1T total | 584 GB | 49.95 t/s UD-Q4_K_XL · 8x MI300X · 10 Aug 2026 The 2-bit cut also runs on the flagship CPU box; that pairing has no published speed. | E64ads_v7ND96is_MI300X_v5 2 deployments›
| $12.02/hr 8x MI300X 1.5 TB | |||||||||||||||||||||||||
Qwen3.8-Max MoE frontier, 2.4T total / 95B active | 1,311 GB | unmeasured No run exists on any box, any engine. Same hardware story as Kimi-K3. | - 1 deployment›
| $11.49/hr | |||||||||||||||||||||||||
Kimi-K3 MoE frontier, 1.5 TB on disk | 1,509 GB | unmeasured Needs 8x MI300X, which Azure could not allocate on spot from our subscriptions when this shipped. On the menu because the box fits it; every claim surface says unmeasured. | - 1 deployment›
| $11.49/hr |
Download sizes are decimal GB, byte-verified against Hugging Face for the quoted variant. All-in prices are the VM spot price as measured in the listed region for our latest snapshot, plus the appliance software fee for that size from the Marketplace price grid. Spot moves; treat these as the shape of the economics, not a quote. Licences are permissive (Apache 2.0, MIT or equivalent) except Kimi, which is modified MIT.
How to read this table
Speeds are single-stream decode through the server. Standard llama.cpp figures use the engine's own default settings; figures labelled ik_llama.cpp use the settings the menu ships for that size. Where a run recorded it, prefill (how fast a prompt is read) is shown after the decode figure. Every figure carries its date because engine versions move. Where a model has been measured on more than one kind of silicon, each figure is listed with its own box rather than averaged. Arm figures are Azure Cobalt, which is CPU-only.
"Runs from" and the measured speed are different boxes
Runs from is the cheapest size we curate that the model actually fits on, with its name next to the price. It is arithmetic against the memory that holds the model (GPU memory, or system memory on a CPU size), with a fifth left over for context, not a speed promise: a 10 GB model does not need an 80 GB card, and pricing it as though it did overstates the cost of starting by four or five times. The measured speed column is usually a bigger, dearer box, because that is where we ran the benchmark. Read them as two facts, not one: the small box will serve the model, more slowly than the figure beside it.
Tested on is a claim about us, not about the model: it lists the boxes we have actually run it on. It is not the list of hardware the model works on, which is far longer. The appliance decides what fits by reading the card at serve time, so a model will happily run on sizes that never appear in that column.
Prices are the cheapest region where our capacity check actually succeeded on 2026-08-25, not the cheapest region Azure lists. Those differ by more than double on some sizes, and a price you cannot get is not a price. Rows marked slice VRAM inferred sit on a partial-card size whose memory we calculated rather than measured, and extra setup means that size needs Secure Boot off and a one-off enable step, so it is cheaper but not a like-for-like deploy.
The appliance menu on the box is authoritative and this page can lag it; use check for new models in the menu for the current list. Hardware, regions and the map are on VM sizes and regions.
Get notified when the catalogue changes
Leave your email and we will tell you when the rankings or the model catalogue change.
Every one of these runs on the same Marketplace image.