Caleta Private AI Appliance / Model catalogue

The measured model catalogue

Every model the appliance ships, with download sizes, measured speeds on real Azure VM sizes, and the deployment pairings we have proven. Measured, not promised.

DEPLOY NOW

Be up and running in minutes: deploy a box into your own subscription, in the region that is cheapest for it today, then serve any model on this page from the appliance menu.

How we measure

Every figure on this page was measured on the exact appliance image we ship, with the serve settings it ships with. Tokens per second are single-stream decode speeds. "Caleta Score" is our in-house infrastructure-as-code benchmark, where higher is better. Deployment pairings marked measured ran on that hardware in that region; estimated pairings are engineering estimates awaiting a measured run.

Deployment pairings and spot prices below are from our latest catalogue snapshot. Spot prices move; the appliance's own spot placement menu action ranks regions from data refreshed daily. The menu on the box is authoritative and this page may lag it.

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.

Pick a model

GLM-5.2 on Azure

Variant	Type	Download	Caleta Score	Measured
GLM-5.2 IQ2_M	MoE, 40B active	239 GB	90	NC96 4xA100: 28 t/s · E64: 7 t/s
GLM-5.2 IQ3_S	MoE, 40B active	288 GB	98	NC96: 26 t/s
GLM-5.2 (744B) (Azure Cobalt Arm)	MoE frontier, 40B active	467 GB	-	~9.5-11.7 t/s Cobalt CPU · frontier flagship; E64ps_v6 only

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
GLM-5.2 (UD-Q4_K_XL)	Standard_ND96is_MI300X_v5	llamacpp	interactive	$11.49 (francecentral)	measured
GLM-5.2 (Q8_0)	Standard_ND96is_MI300X_v5	llamacpp	interactive	$11.49 (francecentral)	measured
GLM-5.2 (UD-IQ2_M)	Standard_NC96ads_A100_v4	llamacpp	interactive	$3.93 (italynorth)	measured
GLM-5.2 (UD-IQ3_S)	Standard_NC96ads_A100_v4	llamacpp	interactive	$3.93 (italynorth)	measured

DeepSeek-R1-0528 on Azure

Variant	Type	Download	Caleta Score	Measured
DeepSeek-R1-0528 (Q2)	MoE, 37B active	234 GB	unscored	E64: 9 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
DeepSeek-R1 (Q8_0)	Standard_ND96is_MI300X_v5	llamacpp	interactive	$11.49 (francecentral)	estimated
DeepSeek-R1 (UD-Q2_K_XL)	Standard_E64ads_v7	llamacpp	batch	$0.98 (eastus)	measured

Devstral-Small-2 on Azure

Variant	Type	Download	Caleta Score	Measured
Devstral-Small-2 (Q4)	dense 24B	14 GB	84	NC80: 92 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Devstral-Small-2 (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Devstral-Small-2 (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Gemma-4-26B-A4B on Azure

Variant	Type	Download	Caleta Score	Measured
Gemma-4-26B-A4B (Q4)	MoE, 4B active	15 GB	82	NC80: 190 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Gemma-4-26B (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Gemma-4-26B (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

GLM-4.7-Flash on Azure

Variant	Type	Download	Caleta Score	Measured
GLM-4.7-Flash (Q4)	MoE, 3B active	18 GB	56	NC80: 170 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
GLM-4.7-Flash (Q4_K_M)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
GLM-4.7-Flash (Q4_K_M)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

gpt-oss-20b on Azure

Variant	Type	Download	Caleta Score	Measured
gpt-oss-20b (MXFP4)	MoE, 4B active	12 GB	60	NC80: 306 t/s
gpt-oss-20b (Azure Cobalt Arm)	MoE reasoner, 4B active	12 GB	-	~75 t/s Cobalt CPU · fastest strong reasoner; fits E8ps_v6

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
gpt-oss-20b (MXFP4)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
gpt-oss-20b (MXFP4)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Kimi-K2.7-Code on Azure

Variant	Type	Download	Caleta Score	Measured
Kimi-K2.7-Code (Q2)	MoE, 32B active	339 GB	unscored	E64: 9.4 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Kimi-K2.7-Code (UD-Q4_K_XL)	Standard_ND96is_MI300X_v5	llamacpp	interactive	$11.49 (francecentral)	measured
Kimi-K2.7-Code (UD-Q2_K_XL)	Standard_E64ads_v7	llamacpp	batch	$0.98 (eastus)	measured

Phi-4 on Azure

Variant	Type	Download	Caleta Score	Measured
Phi-4 (Q4)	dense 14B	9 GB	74	NC80: 152 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Phi-4 (Q4_K_M)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Phi-4 (Q4_K_M)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Qwen2.5-Coder-32B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen2.5-Coder-32B (Q4)	dense 32B	19 GB	86	NC80: 68.4 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen2.5-Coder-32B (Q4_K_M)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Qwen2.5-Coder-32B (Q4_K_M)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Qwen3-235B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3-235B (Q4)	MoE, 22B active	126 GB	86	NC80: 82.5 t/s · E64: 13 t/s
Qwen3-235B-A22B (Azure Cobalt Arm)	MoE big, 22B active	142 GB	-	~15 t/s Cobalt CPU · big brain, RAM-heavy

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3-235B (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured
Qwen3-235B-BF16 (BF16)	Standard_ND96is_MI300X_v5	llamacpp	interactive	$11.49 (francecentral)	measured

Qwen3-Coder-30B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3-Coder-30B (Q4), the trial	MoE, 3B active	18 GB	80	NC80: 263 t/s · E32: 35 t/s
Qwen3-Coder-30B-A3B (Azure Cobalt Arm)	MoE coder, 3B active	19 GB	-	~56-64 t/s Cobalt CPU · fast mature coder; safe default

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3-Coder-30B (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Qwen3-Coder-30B (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Qwen3-Coder-Next-80B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3-Coder-Next-80B (Q4)	MoE, 3B active	46 GB	78	NC80: 169 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3-Coder-Next-80B (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Qwen3-Coder-Next-80B (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Qwen3.6-27B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3.6-27B (Q4)	dense 27B	17 GB	90	NC80 2xH100: 72.7 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3.6-27B (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Qwen3.6-27B (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

Qwen3.6-35B-A3B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3.6-35B-A3B (Q4)	MoE, 3B active	21 GB	76	NC80: 205 t/s
Qwen3.6-35B-A3B (Azure Cobalt Arm)	MoE coder, 3B active	22 GB	-	~34 t/s Cobalt CPU · best coder quality (SWE-bench 73.4)

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3.6-35B (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured
Qwen3.6-35B (UD-Q4_K_XL)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

DeepSeek-V4-Flash-0731 on Azure

Variant	Type	Download	Caleta Score	Measured
DeepSeek-V4-Flash-0731 (MXFP4)	MoE, 13B active	155 GB	unscored	NC80: 54.6 t/s · E64: 9.7 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
DeepSeek-V4-Flash-0731 (MXFP4)	Standard_NC80adis_H100_v5	llamacpp	interactive	$3.64 (indonesiacentral)	measured

gpt-oss-120b on Azure

Variant	Type	Download	Caleta Score	Measured
gpt-oss-120b (MXFP4)	MoE, 5B active	61 GB	unscored	NC24 1xA100: 140 t/s · E64: 48 t/s
gpt-oss-120b (Azure Cobalt Arm)	MoE reasoner, 5B active	63 GB	-	~53 t/s Cobalt CPU · 120B brain at 5B-active speed; value pick

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
gpt-oss-120b (MXFP4)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured

Qwen3-Coder-480B on Azure

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3-Coder-480B (Q8_0)	Standard_ND96is_MI300X_v5	llamacpp	interactive	$11.49 (francecentral)	measured

Qwen3.8-27B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3.8-27B (Q4)	dense 27B	18 GB	unscored	NC24 A100: 47.5 t/s

Proven deployments

Variant	VM size	Engine	Mode	All-in USD/hr*	Status
Qwen3.8-27B (UD-Q4_K_XL)	Standard_NC24ads_A100_v4	llamacpp	interactive	$1.00 (italynorth)	measured

ERNIE-4.5-21B-A3B on Azure

Variant	Type	Download	Caleta Score	Measured
ERNIE-4.5-21B-A3B (Azure Cobalt Arm)	MoE general, 3B active	13 GB	-	~67 t/s Cobalt CPU · tiny, Apache-2.0; fits E8ps_v6

GLM-4.5-Air on Azure

Variant	Type	Download	Caleta Score	Measured
GLM-4.5-Air (Azure Cobalt Arm)	MoE mid, 12B active	73 GB	-	~25 t/s Cobalt CPU · best mid-tier, MIT; needs E16ps_v6 or larger

Qwen3-30B-A3B-Instruct on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3-30B-A3B-Instruct (Azure Cobalt Arm)	MoE general, 3B active	19 GB	-	~63 t/s Cobalt CPU · fast general daily-driver

Qwen3-Next-80B-A3B on Azure

Variant	Type	Download	Caleta Score	Measured
Qwen3-Next-80B-A3B (Azure Cobalt Arm)	MoE general, 3B active	49 GB	-	~28 t/s Cobalt CPU · smart, not a speed pick

*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 as read from the Marketplace price grid. Both land on your Azure bill. Spot moves; treat these as the shape of the economics, not a quote.

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.

Get it on the Azure Marketplace Read the setup guide