harn-vm 0.10.102

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

# ---------- Baseten Model APIs -----------------------------------------------
#
# Baseten serves current open-weight frontier models through an
# OpenAI-compatible /v1/chat/completions surface, advertising
# tools/JSON/structured-output broadly.
#
# GLM used to be pinned to the text grammar here. A 2026-06-23 live probe saw
# GLM-5.2 emit visible `<tool_call><arg_key>...` content instead of OpenAI
# `message.tool_calls`, and that observation was the last survivor of a
# cross-host `glm-5` markup-leak verdict retired on 2026-08-15 (see
# capability_audit.rs). It could not be re-checked at the time because the fleet
# had no Baseten credential, so the row was left conservative rather than
# flipped on another host's evidence.
#
# 2026-08-15 re-probe with a Baseten key: 16/16 clean single
# `message.tool_calls`, across {sync, streaming} x {tool_choice auto, required},
# `finish_reason=tool_calls`, empty assistant content, no markup in any trial.
# The same sweep found GLM-4.7, GLM-5.2-Fast, and every Kimi/DeepSeek/Nemotron/
# GPT-OSS route clean on the native channel, so the pin is retired and Baseten
# now matches every other GLM host. No `glm-5` route anywhere in the catalog
# still carries the markup-leak rationale.

[[provider.baseten]]
model_match = "*glm-5*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
auto_reasoning_overrides = { agent = "off", verify = "off", code = "off" }
prompt_caching = true
honors_chat_template_kwargs = true
chat_template_options_field = "chat_template_args"
reasoning_text_promotable = false
# Baseten reports `supported_sampling_parameters` per route on /v1/models.
# GLM-5.2 and GLM-5.2-Fast list temperature/top_p/stop, so nucleus sampling is
# available but top-k is not. Declaring the absence turns a silently-ignored or
# rejected field into a typed admission error naming the route.
top_p_supported = true
top_k_supported = false
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

[[provider.baseten]]
model_match = "*kimi-k2*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
honors_chat_template_kwargs = true
chat_template_options_field = "chat_template_args"
vision = true
vision_supported = true
# Baseten's Kimi routes list only temperature/stop.
top_p_supported = false
top_k_supported = false
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

[[provider.baseten]]
model_match = "*deepseek-v4*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["effort"]
reasoning_effort_supported = true
reasoning_effort_levels = ["low", "medium", "high"]
prompt_caching = true
# Baseten's DeepSeek V4 routes (undated Pro plus the 0813/0731 snapshots) list
# only temperature/stop.
top_p_supported = false
top_k_supported = false
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "reasoning_summary"

[[provider.baseten]]
model_match = "*gpt-oss*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["effort"]
reasoning_effort_supported = true
reasoning_effort_levels = ["low", "medium", "high"]
reasoning_required_for_tools = true
prompt_caching = true
# Baseten's managed GPT-OSS route lists only temperature/stop. The shared
# `logical:openai-gpt-oss-120b` default still sets `top_p = 1.0` for every
# other host, so 40-defaults/generation.toml keeps `_unset = ["top_p"]` for
# this route; declaring the capability here is what makes that unset checkable
# instead of a lone hand-maintained exception.
top_p_supported = false
top_k_supported = false
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "reasoning_summary"

[[provider.baseten]]
model_match = "*nemotron*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
honors_chat_template_kwargs = true
chat_template_options_field = "chat_template_args"
top_p_supported = true
top_k_supported = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

[[provider.baseten]]
model_match = "*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "none"