Skip to content

Listings

All listings: Novita AI

Every listing for the models on page 4 of 8 of Novita AI, including each model's other versions and routes. Back to Novita AI

Sold by Novita AI

26 listings on this page sold by Novita AI.

Select a score for details.

About these columns
llminfo IQ
llminfo IQ (capability index). How llminfo IQ is computed
Coding IQ
llminfo Coding IQ (coding capability, its own scale). How Coding IQ is computed
Value
Value for money. 50 means as expected for its results. Above 50, cheaper per task. How Value for money works
Yap
Tokens used per coding task vs. models with the same results. Includes agents that write code. Chattiness on maths doesn't predict it. How the Yap score works
Modelllminfo IQCoding IQInput $/1MOutput $/1MMax Input TokensValueYapEpoch Capabilities Index
Kimi K2 0905
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionKimi K3.1We compared the version names in our catalogue.
119.4
From 22 benchmarks. Plenty of data behind it.
≈ 93 ±11
Coding IQ 92.5, 95% interval 82–103. Little data behind it, so treat it as a rough estimate. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.60$2.50262K
56.6
Good value (56.6). Cheaper per task than its benchmark results predict. A fair amount of data behind it.
41 ±9
Terse (41 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
Qwen 3 235B A22B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/qwen/qwen3-235b-a22b-fp8 · fp8
119.3
From 185 benchmarks. Plenty of data behind it.
86 ±12
Coding IQ 86.4, 95% interval 75–98. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.20$0.8041K
53.2
Good value (53.2). Cheaper per task than its benchmark results predict. Plenty of data behind it.
Not measured
143.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
MiMo V2 Flash
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionMiMo V2.6 FlashWe compared release dates and version names in our catalogue.
119.2
From 153 benchmarks. Plenty of data behind it.
95 ±11
Coding IQ 94.7, 95% interval 84–106. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.11cache read $0.024$0.33262K
53.7
Good value (53.7). Cheaper per task than its benchmark results predict. A fair amount of data behind it.
53 ±9
Typical (53 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
Qwen3 Coder
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/qwen/qwen3-coder-480b-a35b-instruct
119.2
From 84 benchmarks. Plenty of data behind it.
93 ±6
Coding IQ 92.6, 95% interval 87–98. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.38$1.55262K
34.9
Poor value (34.9). Costs more per task than its benchmark results predict. Plenty of data behind it.
39 ±9
Terse (39 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
MiniMax M2
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
118.9
From 111 benchmarks. Plenty of data behind it.
95 ±7
Coding IQ 95.4, 95% interval 88–103. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.30cache read $0.03$1.20205K
57.8
Good value (57.8). Cheaper per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.6v
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
118.4
From 45 benchmarks. A fair amount of data behind it.
Not enough data yet$0.30cache read $0.055$0.90131K
44.9
Poor value (44.9). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GPT OSS 120B
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
fp4
118.2
From 424 benchmarks. Plenty of data behind it.
86 ±5
Coding IQ 85.5, 95% interval 80–91. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.05$0.25131K
58.4
Good value (58.4). Cheaper per task than its benchmark results predict. Plenty of data behind it.
37 ±11
Terse (37 ±11). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
140.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Qwen3 Coder Next
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
118.2
From 27 benchmarks. Plenty of data behind it.
98 ±7
Coding IQ 97.7, 95% interval 90–105. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.20$1.50262K
36.8
Poor value (36.8). Costs more per task than its benchmark results predict. Plenty of data behind it.
54 ±9
Typical (54 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
Qwen3 Next 80B A3B Instruct
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
118.1
From 154 benchmarks. Plenty of data behind it.
92 ±11
Coding IQ 91.7, 95% interval 81–102. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.15$1.50131K
47.6
Fair value (47.6). Costs per task about what its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.5 Air
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
bf16
118.0
From 210 benchmarks. Plenty of data behind it.
93 ±7
Coding IQ 93.2, 95% interval 86–100. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.13cache read $0.025$0.85131K
42.6
Poor value (42.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
Qwen3 Coder 30B A3B Instruct
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
116.2
From 18 benchmarks. A fair amount of data behind it.
Not enough data yet$0.07$0.27160K
54.3
Good value (54.3). Cheaper per task than its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measuredNot measured
GLM 4.7 Flash
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionGLM 5.3 FlashWe compared release dates and version names in our catalogue.
bf16
116.1
From 186 benchmarks. Plenty of data behind it.
83 ±10
Coding IQ 82.8, 95% interval 73–93. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.07cache read $0.01$0.40200K
44.1
Poor value (44.1). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.5V
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
115.8
From 62 benchmarks. A fair amount of data behind it.
Not enough data yet$0.60cache read $0.11$1.8066K
36.7
Poor value (36.7). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
Qwen3 32B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/qwen/qwen3-32b-fp8 · fp8
115.6
From 234 benchmarks. Plenty of data behind it.
85 ±9
Coding IQ 84.8, 95% interval 76–93. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.10$0.4541K
35.8
Poor value (35.8). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measured
136.9
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
DeepSeek V3
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Newer version
Newer versionDeepSeek V4We compared the version names in our catalogue.
114.5
From 492 benchmarks. Plenty of data behind it.
85 ±7
Coding IQ 85.3, 95% interval 79–92. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.89$0.8964K
43.6
Poor value (43.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
47 ±9
Typical (47 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
135.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
DeepSeek V3
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Newer version
Newer versionDeepSeek V4We compared the version names in our catalogue.
novita/deepseek/deepseek-v3-0324 · 0324
114.5
From 492 benchmarks. Plenty of data behind it.
85 ±7
Coding IQ 85.3, 95% interval 79–92. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.27cache read $0.135$1.12164K
43.6
Poor value (43.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
47 ±9
Typical (47 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
135.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Qwen3 30B A3B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/qwen/qwen3-30b-a3b-fp8 · fp8
114.4
From 142 benchmarks. Plenty of data behind it.
90 ±10
Coding IQ 90.3, 95% interval 80–101. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.09$0.4541K
35.8
Poor value (35.8). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measured
138.3
Epoch Capabilities IndexSetting: thinkingSource: Epoch AI Benchmarking Hub API
GLM 4 32B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
novita/thudm/glm-4-32b-0414 · 0414
113.3
From 10 benchmarks. A fair amount of data behind it.
Not enough data yet$0.55$1.6632K
58.4
Good value (58.4). Cheaper per task than its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measuredNot measured
Qwen3 8B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/qwen/qwen3-8b-fp8 · fp8
112.4
From 182 benchmarks. Plenty of data behind it.
80 ±12
Coding IQ 80.3, 95% interval 68–92. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.035$0.138128K
58.0
Good value (58.0). Cheaper per task than its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measured
132.1
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Llama 4 Maverick
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/meta-llama/llama-4-maverick-17b-128e-instruct-fp8 · fp8
111.8
From 350 benchmarks. Plenty of data behind it.
78 ±9
Coding IQ 78.1, 95% interval 69–87. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.27$0.851M
42.1
Poor value (42.1). Costs more per task than its benchmark results predict. Plenty of data behind it.
38 ±9
Terse (38 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
132.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
qwen3-vl-30b
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/qwen/qwen3-vl-30b-a3b-thinking · thinking
111.5
From 81 benchmarks. Plenty of data behind it.
Not enough data yet$0.20$1.00131K
45.6
Poor value (45.6). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
qwen3-vl-30b
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
qwen/qwen3-vl-30b-a3b-instruct
111.5
From 81 benchmarks. Plenty of data behind it.
Not enough data yet$0.20$0.70131K
45.6
Poor value (45.6). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GPT OSS 20b
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
fp4
111.3
From 251 benchmarks. Plenty of data behind it.
88 ±12
Coding IQ 88.2, 95% interval 76–101. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.04$0.15131K
48.9
Fair value (48.9). Costs per task about what its benchmark results predict. Plenty of data behind it.
40 ±9
Terse (40 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
134.9
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Qwen3 Embedding 0.6B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
110.5
From 192 benchmarks. Plenty of data behind it.
Not enough data yet$0.07—33KNot enough data yetNot measuredNot measured
DeepSeek R1 0528 Qwen3 8B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
≈ 110
From 9 benchmarks. Little data behind it, so treat it as a rough estimate.
Not enough data yet$0.06$0.09128K
43.4
Poor value (43.4). Costs more per task than its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measuredNot measured
Nemotron 3 Nano 30B A3B
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
109.7
From 137 benchmarks. Plenty of data behind it.
91 ±14
Coding IQ 91.4, 95% interval 77–106. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.05$0.20262K
36.7
Poor value (36.7). Costs more per task than its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measuredNot measured

Run by Novita AI

36 listings on this page where Novita AI runs the model. Another provider sells these listings.

Select a score for details.

About these columns
llminfo IQ
llminfo IQ (capability index). How llminfo IQ is computed
Coding IQ
llminfo Coding IQ (coding capability, its own scale). How Coding IQ is computed
Value
Value for money. 50 means as expected for its results. Above 50, cheaper per task. How Value for money works
Yap
Tokens used per coding task vs. models with the same results. Includes agents that write code. Chattiness on maths doesn't predict it. How the Yap score works
ModelProviderllminfo IQCoding IQInput $/1MOutput $/1MMax Input TokensValueYapEpoch Capabilities Index
Kimi K2 0905
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionKimi K3.1We compared the version names in our catalogue.
fp8
 OpenRouter
119.4
From 22 benchmarks. Plenty of data behind it.
≈ 93 ±11
Coding IQ 92.5, 95% interval 82–103. Little data behind it, so treat it as a rough estimate. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.60$2.50262K
56.6
Good value (56.6). Cheaper per task than its benchmark results predict. A fair amount of data behind it.
41 ±9
Terse (41 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
Qwen 3 235B A22B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
119.3
From 185 benchmarks. Plenty of data behind it.
86 ±12
Coding IQ 86.4, 95% interval 75–98. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.20$0.8041K
53.2
Good value (53.2). Cheaper per task than its benchmark results predict. Plenty of data behind it.
Not measured
143.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
MiMo V2 Flash
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionMiMo V2.6 FlashWe compared release dates and version names in our catalogue.
 Hugging Face
119.2
From 153 benchmarks. Plenty of data behind it.
95 ±11
Coding IQ 94.7, 95% interval 84–106. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.10$0.30262K
53.7
Good value (53.7). Cheaper per task than its benchmark results predict. A fair amount of data behind it.
53 ±9
Typical (53 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
Qwen3 Coder
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Qwen/Qwen3-Coder-480B-A35B-Instruct
 Hugging Face
119.2
From 84 benchmarks. Plenty of data behind it.
93 ±6
Coding IQ 92.6, 95% interval 87–98. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.30$1.30262K
34.9
Poor value (34.9). Costs more per task than its benchmark results predict. Plenty of data behind it.
39 ±9
Terse (39 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
MiniMax M2
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
118.9
From 111 benchmarks. Plenty of data behind it.
95 ±7
Coding IQ 95.4, 95% interval 88–103. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.30$1.20205K
57.8
Good value (57.8). Cheaper per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.6v
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Llmgateway
118.4
From 45 benchmarks. A fair amount of data behind it.
Not enough data yet$0.30cache read $0.055$0.90Not published by the provider
44.9
Poor value (44.9). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.6v
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
bf16
 OpenRouter
118.4
From 45 benchmarks. A fair amount of data behind it.
Not enough data yet$0.30cache read $0.055$0.90131K
44.9
Poor value (44.9). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GPT OSS 120B
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
118.2
From 424 benchmarks. Plenty of data behind it.
86 ±5
Coding IQ 85.5, 95% interval 80–91. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.05$0.25131K
58.4
Good value (58.4). Cheaper per task than its benchmark results predict. Plenty of data behind it.
37 ±11
Terse (37 ±11). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
140.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Qwen3 Coder Next
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
118.2
From 27 benchmarks. Plenty of data behind it.
98 ±7
Coding IQ 97.7, 95% interval 90–105. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.20$1.50262K
36.8
Poor value (36.8). Costs more per task than its benchmark results predict. Plenty of data behind it.
54 ±9
Typical (54 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
Not measured
Qwen3 Next 80B A3B Instruct
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
118.1
From 154 benchmarks. Plenty of data behind it.
92 ±11
Coding IQ 91.7, 95% interval 81–102. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.15$1.50131K
47.6
Fair value (47.6). Costs per task about what its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.5 Air
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
118.0
From 210 benchmarks. Plenty of data behind it.
93 ±7
Coding IQ 93.2, 95% interval 86–100. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.13$0.85131K
42.6
Poor value (42.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
GLM 4.5 Air
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
bf16 · 🇺🇸 US
 NanoGPT
118.0
From 210 benchmarks. Plenty of data behind it.
93 ±7
Coding IQ 93.2, 95% interval 86–100. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.1365cache read $0.0263$0.8925Not published by the provider
42.6
Poor value (42.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
GLM 4.5 Air
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
z-ai/GLM-4.5-Air:thinking · thinking · bf16 · 🇺🇸 US
 NanoGPT
118.0
From 210 benchmarks. Plenty of data behind it.
93 ±7
Coding IQ 93.2, 95% interval 86–100. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.1365cache read $0.0263$0.8925Not published by the provider
42.6
Poor value (42.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
GLM 4.5 Air
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
bf16
 OpenRouter
118.0
From 210 benchmarks. Plenty of data behind it.
93 ±7
Coding IQ 93.2, 95% interval 86–100. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.13cache read $0.025$0.85131K
42.6
Poor value (42.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
Kimi K2
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionKimi K3.1We compared the version names in our catalogue.
moonshotai/kimi-k2-instruct
 Llmgateway
117.4
From 30 benchmarks. Plenty of data behind it.
≈ 95 ±16stale
Coding IQ 95.1, 95% interval 80–111. Little data behind it, so treat it as a rough estimate. Uses the model's best published results. It has its own scale, separate from llminfo IQ. Its newest result is over a year old.
$0.57$2.30Not published by the provider
52.3
Fair value (52.3). Costs per task about what its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measured
141.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Kimi K2
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionKimi K3.1We compared the version names in our catalogue.
fp8
 OpenRouter
117.4
From 30 benchmarks. Plenty of data behind it.
≈ 95 ±16stale
Coding IQ 95.1, 95% interval 80–111. Little data behind it, so treat it as a rough estimate. Uses the model's best published results. It has its own scale, separate from llminfo IQ. Its newest result is over a year old.
$0.57$2.30131K
52.3
Fair value (52.3). Costs per task about what its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measured
141.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
GLM 4.7 Flash
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionGLM 5.3 FlashWe compared release dates and version names in our catalogue.
 Hugging Face
116.1
From 186 benchmarks. Plenty of data behind it.
83 ±10
Coding IQ 82.8, 95% interval 73–93. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.07$0.40200K
44.1
Poor value (44.1). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.7 Flash
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionGLM 5.3 FlashWe compared release dates and version names in our catalogue.
bf16 · 🇺🇸 US
 NanoGPT
116.1
From 186 benchmarks. Plenty of data behind it.
83 ±10
Coding IQ 82.8, 95% interval 73–93. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.0735cache read $0.0105$0.42Not published by the provider
44.1
Poor value (44.1). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.7 Flash
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Newer version
Newer versionGLM 5.3 FlashWe compared release dates and version names in our catalogue.
z-ai/glm-4.7-flash:thinking · thinking · bf16 · 🇺🇸 US
 NanoGPT
116.1
From 186 benchmarks. Plenty of data behind it.
83 ±10
Coding IQ 82.8, 95% interval 73–93. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.0735cache read $0.0105$0.42Not published by the provider
44.1
Poor value (44.1). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GLM 4.5V
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
115.8
From 62 benchmarks. A fair amount of data behind it.
Not enough data yet$0.60$1.8066K
36.7
Poor value (36.7). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
GLM 4.5V
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Llmgateway
115.8
From 62 benchmarks. A fair amount of data behind it.
Not enough data yet$0.60cache read $0.11$1.80Not published by the provider
36.7
Poor value (36.7). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
GLM 4.5V
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
fp8
 OpenRouter
115.8
From 62 benchmarks. A fair amount of data behind it.
Not enough data yet$0.60cache read $0.11$1.8066K
36.7
Poor value (36.7). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measuredNot measured
Qwen3 32B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
115.6
From 234 benchmarks. Plenty of data behind it.
85 ±9
Coding IQ 84.8, 95% interval 76–93. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.10$0.4541K
35.8
Poor value (35.8). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measured
136.9
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
DeepSeek V3
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Newer version
Newer versionDeepSeek V4We compared the version names in our catalogue.
 Hugging Face
114.5
From 492 benchmarks. Plenty of data behind it.
85 ±7
Coding IQ 85.3, 95% interval 79–92. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.40$1.3064K
43.6
Poor value (43.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
47 ±9
Typical (47 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
135.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
DeepSeek V3
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Newer version
Newer versionDeepSeek V4We compared the version names in our catalogue.
deepseek-ai/DeepSeek-V3-0324 · 0324
 Hugging Face
114.5
From 492 benchmarks. Plenty of data behind it.
85 ±7
Coding IQ 85.3, 95% interval 79–92. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.27$1.12164K
43.6
Poor value (43.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
47 ±9
Typical (47 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
135.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
DeepSeek V3
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Newer version
Newer versionDeepSeek V4We compared the version names in our catalogue.
 Requesty
114.5
From 492 benchmarks. Plenty of data behind it.
85 ±7
Coding IQ 85.3, 95% interval 79–92. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.89$0.8964K
43.6
Poor value (43.6). Costs more per task than its benchmark results predict. Plenty of data behind it.
47 ±9
Typical (47 ±9). Generates about as many tokens as models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
135.3
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Qwen3 30B A3B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
114.4
From 142 benchmarks. Plenty of data behind it.
90 ±10
Coding IQ 90.3, 95% interval 80–101. A fair amount of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.09$0.4541K
35.8
Poor value (35.8). Costs more per task than its benchmark results predict. Plenty of data behind it.
Not measured
138.3
Epoch Capabilities IndexSetting: thinkingSource: Epoch AI Benchmarking Hub API
GLM 4 32B
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
zai-org/GLM-4-32B-0414 · 0414
 Hugging Face
113.3
From 10 benchmarks. A fair amount of data behind it.
Not enough data yet$0.55$1.6632K
58.4
Good value (58.4). Cheaper per task than its benchmark results predict. Little data behind it, so treat it as a rough estimate.
Not measuredNot measured
Llama 4 Maverick
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 · fp8
 Hugging Face
111.8
From 350 benchmarks. Plenty of data behind it.
78 ±9
Coding IQ 78.1, 95% interval 69–87. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.27$0.851M
42.1
Poor value (42.1). Costs more per task than its benchmark results predict. Plenty of data behind it.
38 ±9
Terse (38 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
132.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Llama 4 Maverick
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
meta-llama/llama-4-maverick-17b-128e-instruct-fp8
 Llmgateway
111.8
From 350 benchmarks. Plenty of data behind it.
78 ±9
Coding IQ 78.1, 95% interval 69–87. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.27$0.85Not published by the provider
42.1
Poor value (42.1). Costs more per task than its benchmark results predict. Plenty of data behind it.
38 ±9
Terse (38 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
132.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
Llama 4 Maverick
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
novita/meta-llama/llama-4-maverick-17b-128e-instruct-fp8 · fp8
 Requesty
111.8
From 350 benchmarks. Plenty of data behind it.
78 ±9
Coding IQ 78.1, 95% interval 69–87. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.20$0.851M
42.1
Poor value (42.1). Costs more per task than its benchmark results predict. Plenty of data behind it.
38 ±9
Terse (38 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
132.2
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API
qwen3-vl-30b
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Qwen/Qwen3-VL-30B-A3B-Thinking · thinking
 Hugging Face
111.5
From 81 benchmarks. Plenty of data behind it.
Not enough data yet$0.20$1.00131K
45.6
Poor value (45.6). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
qwen3-vl-30b
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
Qwen/Qwen3-VL-30B-A3B-Instruct
 Hugging Face
111.5
From 81 benchmarks. Plenty of data behind it.
Not enough data yet$0.20$0.70131K
45.6
Poor value (45.6). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
qwen3-vl-30b
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
qwen/qwen3-vl-30b-a3b-instruct
 Llmgateway
111.5
From 81 benchmarks. Plenty of data behind it.
Not enough data yet$0.20$0.70Not published by the provider
45.6
Poor value (45.6). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
qwen3-vl-30b
🇨🇳
Maker headquartered in CN, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
qwen/qwen3-vl-30b-a3b-instruct · bf16
 OpenRouter
111.5
From 81 benchmarks. Plenty of data behind it.
Not enough data yet$0.20$0.70131K
45.6
Poor value (45.6). Costs more per task than its benchmark results predict. A fair amount of data behind it.
Not measuredNot measured
GPT OSS 20b
🇺🇸
Maker headquartered in US, according to the company's own website. The citation is on the model page.
Open weights. The weights are publicly available, according to the provider's model list or a public Hugging Face repo.
 Hugging Face
111.3
From 251 benchmarks. Plenty of data behind it.
88 ±12
Coding IQ 88.2, 95% interval 76–101. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.
$0.04$0.15131K
48.9
Fair value (48.9). Costs per task about what its benchmark results predict. Plenty of data behind it.
40 ±9
Terse (40 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.
134.9
Epoch Capabilities IndexSource: Epoch AI Benchmarking Hub API

Data credits: Epoch AI Benchmarking Hub, CC BY 4.0 (licence)

How this data works