What the model costs, and where it can run.

Pricing & Capabilities lists every tracked model. Compare places three models side by side. Compliance & Residency cites HIPAA, FedRAMP, zero data retention, and EU residency for each cloud host. On Hardware, FP16, FP8, and INT4 are weights only, with a total when KV is computed. KV is 16-bit unless the card states another format.

Weekly on Mondays at 4:30 AM PT. Next refresh Mon 26 Oct.

Index accurately reflects market data as of 26 Oct 2026.

This week

Open-weight memory uses the model card or Hugging Face config. An estimate is labeled.
39 published models from 11 companies

Scroll sideways to see quantization columns.

Open-weight models only. FP16, FP8, and INT4 are weights only. Where KV is computed, the cell states the total with that cache. KV is 16-bit unless the card states another format. Checkpoint, minimum GPU, and vLLM follow the cited page.

Model	Parameters	FP16 weights	FP8 weights	INT4 weights	KV	Checkpoint	Minimum GPU	vLLM
Command A+CohereModels Model page BF16 FP8 W4A4 vLLM	25B active / 218B total	406.1 GiBWeights only. Total with KV 407.4 GiB. Estimate. Hypothetical uniform weights.	203.0 GiBWeights only. Total with KV 204.4 GiB. Estimate. Hypothetical uniform weights.	101.5 GiBWeights only. Total with KV 102.9 GiB. Estimate. Hypothetical uniform weights.	1.38 GiBEstimate. 16-bit KV. Full layers at 32,768 tokens, sliding layers at their window.	BF16, FP8, and W4A4	1× B200 at W4A4 or 2× H100 at W4A4	Listed
DeepSeek V4 ProDeepSeekModel card vLLM	49B active / 1.6T total	2980.2 GiBWeights only. Estimate. Hypothetical uniform weights.	1490.1 GiBWeights only. Estimate. Hypothetical uniform weights.	745.1 GiBWeights only. Estimate. Hypothetical uniform weights.	Not published	FP4 + FP8 mixed. Experts are FP4. Most other parameters are FP8.	Not published	Listed
DeepSeek V4.1 FlashDeepSeekModel card Config	8B/16B active / 552B backbone + 196B Engram	1393.3 GiBWeights only. Total with KV 1393.3 GiB. Estimate. Hypothetical uniform weights for 552B backbone + 196B Engram. 196B of the 748B are Engram parameters, sparsely accessed via token-based lookup, and are still counted as stored weights.	696.6 GiBWeights only. Total with KV 696.7 GiB. Estimate. Hypothetical uniform weights for 552B backbone + 196B Engram. 196B of the 748B are Engram parameters, sparsely accessed via token-based lookup, and are still counted as stored weights.	348.3 GiBWeights only. Total with KV 348.3 GiB. Estimate. Hypothetical uniform weights for 552B backbone + 196B Engram. 196B of the 748B are Engram parameters, sparsely accessed via token-based lookup, and are still counted as stored weights.	0.03 GiBFP4 main KV cache from the model card: 890 bytes per token at 32,768 tokens. Exception to the 16-bit KV assumption.	fp8 weights. Expert dtype is fp4.	Not published	Not listed
Gemma 4 26BGoogleModel card Config vLLM	3.8B active / 25.2B total. Hugging Face model card (26B-A4B).	46.9 GiBWeights only. Total with KV 47.8 GiB. Estimate. Hypothetical uniform weights.	23.5 GiBWeights only. Total with KV 24.3 GiB. Estimate. Hypothetical uniform weights.	11.7 GiBWeights only. Total with KV 12.6 GiB. Estimate. Hypothetical uniform weights.	0.82 GiBEstimate. 16-bit KV. Full layers at 32,768 tokens, sliding layers at their window. Full layers use the config's global head size and global KV heads. attention_k_eq_v is true, so those layers have no separate V projection, but K and V are both stored after separate norms and num_kv_shared_layers is 0, so full-layer KV is not halved.	bfloat16	Not published	Architecture supported
Gemma 4 31BGoogleModel card Config vLLM	30.7B. Hugging Face model card.	57.2 GiBWeights only. Total with KV 60.5 GiB. Estimate. Hypothetical uniform weights.	28.6 GiBWeights only. Total with KV 31.9 GiB. Estimate. Hypothetical uniform weights.	14.3 GiBWeights only. Total with KV 17.6 GiB. Estimate. Hypothetical uniform weights.	3.28 GiBEstimate. 16-bit KV. Full layers at 32,768 tokens, sliding layers at their window. Full layers use the config's global head size and global KV heads. attention_k_eq_v is true, so those layers have no separate V projection, but K and V are both stored after separate norms and num_kv_shared_layers is 0, so full-layer KV is not halved.	bfloat16	Not published	Architecture supported
gpt-oss-120bOpenAIModel page Config vLLM	5.1B active / 117B total	217.9 GiBWeights only. Total with KV 219.1 GiB. Estimate. Hypothetical uniform weights.	109.0 GiBWeights only. Total with KV 110.1 GiB. Estimate. Hypothetical uniform weights.	54.5 GiBWeights only. Total with KV 55.6 GiB. Estimate. Hypothetical uniform weights.	1.13 GiBEstimate. 16-bit KV. Full layers at 32,768 tokens, sliding layers at their window.	mxfp4. Attention, the router, embeddings, and the output layer are not quantized.	Not published	Listed
gpt-oss-20bOpenAIModel page Config vLLM	3.6B active / 21B total	39.1 GiBWeights only. Total with KV 39.9 GiB. Estimate. Hypothetical uniform weights.	19.6 GiBWeights only. Total with KV 20.3 GiB. Estimate. Hypothetical uniform weights.	9.78 GiBWeights only. Total with KV 10.5 GiB. Estimate. Hypothetical uniform weights.	0.75 GiBEstimate. 16-bit KV. Full layers at 32,768 tokens, sliding layers at their window.	mxfp4. Attention, the router, embeddings, and the output layer are not quantized.	Not published	Listed
Llama 4 MaverickMetaModel card vLLM	17B active / 400B total	745.1 GiBWeights only. Estimate. Hypothetical uniform weights.	372.5 GiBWeights only. Estimate. Hypothetical uniform weights.	186.3 GiBWeights only. Estimate. Hypothetical uniform weights.	Not published	BF16 and FP8	FP8: one H100 DGX host	Listed
Llama 4 ScoutMetaModel card vLLM	17B active / 109B total	203.0 GiBWeights only. Estimate. Hypothetical uniform weights.	101.5 GiBWeights only. Estimate. Hypothetical uniform weights.	50.8 GiBWeights only. Estimate. Hypothetical uniform weights.	Not published	BF16	One H100 at INT4	Listed
Mistral Large 4MistralModel card	52B active / 1.05T total	1955.8 GiBWeights only. Estimate. Hypothetical uniform weights. The weights are not posted yet.	977.9 GiBWeights only. Estimate. Hypothetical uniform weights. The weights are not posted yet.	488.9 GiBWeights only. Estimate. Hypothetical uniform weights. The weights are not posted yet.	Not published	Weights are not posted yet.	Not published	Not listed
Mistral Medium 3.5MistralModel card Config vLLM	128B dense	238.4 GiBWeights only. Total with KV 249.4 GiB. Estimate. Hypothetical uniform weights.	119.2 GiBWeights only. Total with KV 130.2 GiB. Estimate. Hypothetical uniform weights.	59.6 GiBWeights only. Total with KV 70.6 GiB. Estimate. Hypothetical uniform weights.	11.0 GiBEstimate. 16-bit KV.	FP8 (5.9B modules left in BF16).	Not published	Architecture supported
Mistral Small 4MistralModel page Config vLLM	6.5B active / 119B total	221.7 GiBWeights only. Total with KV 222.4 GiB. Estimate. Hypothetical uniform weights.	110.8 GiBWeights only. Total with KV 111.5 GiB. Estimate. Hypothetical uniform weights.	55.4 GiBWeights only. Total with KV 56.1 GiB. Estimate. Hypothetical uniform weights.	0.70 GiBEstimate. 16-bit KV.	fp8. The vision tower, projector, and output layer are not quantized.	Not published	Architecture supported
Phi-4-reasoning-vision-15BMicrosoftModel card Config vLLM	15B	27.9 GiBWeights only. Total with KV 31.1 GiB. Estimate. Hypothetical uniform weights.	14.0 GiBWeights only. Total with KV 17.1 GiB. Estimate. Hypothetical uniform weights.	6.98 GiBWeights only. Total with KV 10.1 GiB. Estimate. Hypothetical uniform weights.	3.13 GiBEstimate. 16-bit KV.	bfloat16	Not published	Listed

FP16, FP8, and INT4 are hypothetical uniform-precision weights: the published total parameter count times 2 bytes, 1 byte, or 0.5 bytes, with 1 GiB = 2^30 bytes. Those columns are weights only. Where KV is computed, each cell also states the total with that cache. A mixture-of-experts total includes every expert. When a card publishes more than one total, the cell names the total used. KV is separate and assumed 16-bit: 2 × layers × key-value heads × head dimension × tokens × 2 bytes, or for multi-latent attention layers × (kv_lora_rank + rope head dimension) × tokens × 2 bytes, unless the card states another KV format. Full layers use 32,768 tokens, or the published context when that is shorter. Where the config publishes layer_types and a sliding window, sliding layers use that window. Gemma full layers use global_head_dim and num_global_key_value_heads; its sliding layers use head_dim and num_key_value_heads. Gemma attention_k_eq_v drops the V projection on full layers, but the cache still stores K and V after separate norms, and num_kv_shared_layers is 0, so that KV is not halved. KV is left unpublished only when that layout cannot be derived. No framework overhead is added. Checkpoint precision is the format the repository publishes. Minimum GPU is shown only when the card states a device. vLLM is Listed when its supported-models page names the checkpoint, Architecture supported when that page names the architecture but not this checkpoint, and Not listed otherwise.

Loading the index.