harn-vm 0.10.101

Async bytecode virtual machine for the Harn programming language
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

# Baseten Model APIs. Live /v1/models re-verified 2026-08-15 with
# BASETEN_API_KEY: ids, served context windows, pricing, features, modalities,
# and route quantization come from the provider-owned endpoint. Rows are keyed
# with `baseten/<wire-id>` so failover/eval tooling can compare the same
# weights across Baseten, Z.AI, Together, DeepInfra, OpenRouter, and NVIDIA.
#
# The 2026-08-15 sweep POSTed a 5-token completion to every row rather than
# diffing the models listing, because a listed id is not necessarily a callable
# one. Four rows returned HTTP 410 "the model version you are trying to access
# has been deprecated" and were removed:
#
#   zai-org/GLM-5, zai-org/GLM-5.1  -- superseded by GLM-5.2 / GLM-5.2-Fast
#   moonshotai/Kimi-K2.5            -- superseded by Kimi-K2.6 / K2.7-Code
#   nvidia/Nemotron-120B-A12B       -- no successor at that size is served
#
# The same sweep found two context windows and one cache-read rate stale
# against the provider-owned metadata (GLM-5.2 was catalogued at 262144 tokens
# and $0.26/Mtok cache reads against a served 1048576 and $0.14; DeepSeek V4
# Pro at 131000 against a served 262144), so every surviving row below now
# carries the value the endpoint reports.

[models."baseten/zai-org/GLM-5.2"]
name = "GLM 5.2 (Baseten)"
provider = "baseten"
wire_model = "zai-org/GLM-5.2"
logical_model = "glm-5.2"
equivalence_group = "glm-5.2"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 1048576
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.40, output_per_mtok = 4.40, cache_read_per_mtok = 0.14 }
architecture = { parameter_count_b = 744.0, active_parameter_count_b = 40.0, moe = true, quantization = "fp8", license = "MIT", source_url = "https://www.baseten.co/blog/how-we-built-the-worlds-fastest-api-for-glm-52/", last_verified = "2026-08-15" }
performance = { output_tokens_per_sec = 281.2, time_to_answer_s = 7.95, source = "artificial_analysis", source_url = "https://artificialanalysis.ai/models/glm-5-2/providers", last_verified = "2026-06-23", notes = "Artificial Analysis GLM-5.2 provider benchmark. The 2026-06-23 Harn probe that accompanied this figure also recorded native-tool unreliability; a 2026-08-15 re-probe found 16/16 clean native tool calls and that verdict was withdrawn." }
tier = "frontier"
open_weight = true
strengths = ["speed", "coding", "agentic", "tool_use", "reasoning", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "kimi", "deepseek"]

# Same weights as the row above (Baseten reports `hugging_face_id
# zai-org/GLM-5.2` for both); the `-Fast` id is a higher-throughput serving
# tier priced at a 50% premium. Modelled the way `deepinfra-turbo` and
# `moonshot-direct-highspeed` are: one equivalence group, distinct
# served_variant.
[models."baseten/zai-org/GLM-5.2-Fast"]
name = "GLM 5.2 Fast (Baseten)"
provider = "baseten"
wire_model = "zai-org/GLM-5.2-Fast"
logical_model = "glm-5.2"
equivalence_group = "glm-5.2"
served_variant = "baseten-model-api-fast"
api_dialect = "openai_chat"
context_window = 1048576
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 2.10, output_per_mtok = 6.60, cache_read_per_mtok = 0.21 }
architecture = { parameter_count_b = 744.0, active_parameter_count_b = 40.0, moe = true, quantization = "fp8", license = "MIT", source_url = "https://www.baseten.co/blog/how-we-built-the-worlds-fastest-api-for-glm-52/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["speed", "coding", "agentic", "tool_use", "reasoning", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "kimi", "deepseek"]

[models."baseten/moonshotai/Kimi-K2.7-Code"]
name = "Kimi K2.7 Code (Baseten)"
provider = "baseten"
wire_model = "moonshotai/Kimi-K2.7-Code"
logical_model = "moonshot-kimi-k2.7-code"
equivalence_group = "moonshot-kimi-k2.7-code"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 262000
capabilities = ["tools", "vision", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.95, output_per_mtok = 4.00, cache_read_per_mtok = 0.16 }
architecture = { quantization = "fp4", source_url = "https://www.baseten.co/library/kimi-k27-code/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "tool_use", "reasoning", "long_context", "vision"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]

[models."baseten/deepseek-ai/DeepSeek-V4-Pro"]
name = "DeepSeek V4 Pro (Baseten)"
provider = "baseten"
wire_model = "deepseek-ai/DeepSeek-V4-Pro"
logical_model = "deepseek-v4-pro"
equivalence_group = "deepseek-v4-pro"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 262144
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.74, output_per_mtok = 3.48, cache_read_per_mtok = 0.145 }
architecture = { parameter_count_b = 1600.0, active_parameter_count_b = 49.0, moe = true, quantization = "fp4", source_url = "https://www.baseten.co/library/deepseek-v4/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["reasoning", "coding", "agentic", "tool_use", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "kimi"]

# Dated snapshots. Baseten serves these alongside the undated ids, and they are
# the builds an eval should pin so a silent upstream reroute cannot change what
# a recorded trial measured.
[models."baseten/deepseek-ai/DeepSeek-V4-Pro-0813"]
name = "DeepSeek V4 Pro 0813 (Baseten)"
provider = "baseten"
wire_model = "deepseek-ai/DeepSeek-V4-Pro-0813"
logical_model = "deepseek-v4-pro-0813"
equivalence_group = "deepseek-v4-pro-0813"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 1048576
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.32, output_per_mtok = 3.96, cache_read_per_mtok = 0.132 }
architecture = { parameter_count_b = 1600.0, active_parameter_count_b = 49.0, moe = true, quantization = "fp4", source_url = "https://www.baseten.co/library/deepseek-v4/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["reasoning", "coding", "agentic", "tool_use", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "kimi"]

[models."baseten/deepseek-ai/DeepSeek-V4-Flash-0731"]
name = "DeepSeek V4 Flash 0731 (Baseten)"
provider = "baseten"
wire_model = "deepseek-ai/DeepSeek-V4-Flash-0731"
logical_model = "deepseek-v4-flash-0731"
equivalence_group = "deepseek-v4-flash-0731"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 1048576
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.13, output_per_mtok = 0.26, cache_read_per_mtok = 0.028 }
architecture = { parameter_count_b = 284.0, active_parameter_count_b = 13.0, moe = true, quantization = "fp8", source_url = "https://www.baseten.co/library/deepseek-v4/", last_verified = "2026-08-15" }
tier = "mid"
open_weight = true
strengths = ["speed", "cheap", "reasoning", "coding", "agentic", "tool_use", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "kimi"]

[models."baseten/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B"]
name = "Nemotron 3 Ultra 550B A55B (Baseten)"
provider = "baseten"
wire_model = "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B"
logical_model = "nemotron-3-ultra-550b-a55b"
equivalence_group = "nemotron-3-ultra-550b-a55b"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 202800
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.60, output_per_mtok = 2.40, cache_read_per_mtok = 0.12 }
architecture = { parameter_count_b = 550.0, active_parameter_count_b = 55.0, moe = true, quantization = "fp4", source_url = "https://www.baseten.co/library/nemotron-3-ultra/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["reasoning", "coding", "agentic", "tool_use", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek", "kimi"]

[models."baseten/openai/gpt-oss-120b"]
name = "GPT-OSS 120B (Baseten)"
provider = "baseten"
wire_model = "openai/gpt-oss-120b"
logical_model = "openai-gpt-oss-120b"
equivalence_group = "openai-gpt-oss-120b"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 128072
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.10, output_per_mtok = 0.50, cache_read_per_mtok = 0.10 }
architecture = { parameter_count_b = 117.0, active_parameter_count_b = 5.1, moe = true, quantization = "fp4", license = "Apache-2.0", source_url = "https://www.baseten.co/library/gpt-oss-120b/", last_verified = "2026-08-15" }
tier = "mid"
open_weight = true
strengths = ["cheap", "tool_use", "reasoning", "coding"]
complementary_with = ["anthropic-claude", "google-gemini", "qwen", "deepseek", "kimi"]

[models."baseten/zai-org/GLM-4.7"]
name = "GLM 4.7 (Baseten)"
provider = "baseten"
wire_model = "zai-org/GLM-4.7"
logical_model = "glm-4.7"
equivalence_group = "glm-4.7"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching"]
pricing = { input_per_mtok = 0.60, output_per_mtok = 2.20, cache_read_per_mtok = 0.12 }
architecture = { quantization = "fp4", source_url = "https://www.baseten.co/pricing/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "tool_use", "long_context"]

[models."baseten/moonshotai/Kimi-K2.6"]
name = "Kimi K2.6 (Baseten)"
provider = "baseten"
wire_model = "moonshotai/Kimi-K2.6"
logical_model = "moonshot-kimi-k2.6"
equivalence_group = "moonshot-kimi-k2.6"
served_variant = "baseten-model-api"
api_dialect = "openai_chat"
context_window = 262000
capabilities = ["tools", "vision", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.95, output_per_mtok = 4.00, cache_read_per_mtok = 0.16 }
architecture = { quantization = "fp4", source_url = "https://www.baseten.co/pricing/", last_verified = "2026-08-15" }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "tool_use", "reasoning", "long_context", "vision"]