harn-vm 0.9.21

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

# Current Together serverless frontier/agent routes observed through
# `GET /v1/models`. Keep rows here when Harn supports the provider but the
# model is absent from direct-provider catalogs.

[models."Qwen/Qwen3.7-Max"]
name = "Qwen3.7 Max (Together)"
provider = "together"
context_window = 1000000
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.25, output_per_mtok = 3.75, cache_read_per_mtok = 0.125 }

tier = "frontier"
open_weight = true
strengths = ["reasoning", "coding", "tool_use", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "kimi"]
[models."deepseek-ai/DeepSeek-V4-Pro"]
name = "DeepSeek V4 Pro (Together)"
provider = "together"
context_window = 512000
equivalence_group = "frontier-agent-coding"
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.74, output_per_mtok = 3.48, cache_read_per_mtok = 0.20 }

tier = "frontier"
open_weight = true
strengths = ["agentic", "coding", "long_context", "tool_use", "reasoning"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]
[models."zai-org/GLM-5.1"]
name = "GLM 5.1 (Together)"
provider = "together"
context_window = 202752
equivalence_group = "frontier-agent-coding"
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.40, output_per_mtok = 4.40, cache_read_per_mtok = 0.26 }

tier = "frontier"
open_weight = true
strengths = ["agentic", "coding", "long_context", "tool_use", "reasoning"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]
[models."zai-org/GLM-5.2"]
name = "GLM 5.2 (Together)"
provider = "together"
# Together serves GLM-5.2 at 262K (fp4), not the model's full 1M window
# — per-host context caps differ; verified 2026-07-02 on together.ai.
context_window = 262144
equivalence_group = "frontier-agent-coding"
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 1.40, output_per_mtok = 4.40, cache_read_per_mtok = 0.26 }
tier = "frontier"
open_weight = true
strengths = ["agentic", "coding", "long_context", "tool_use", "reasoning"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]

# MiniMax M3 on Together — 524K serverless window, billed at the same
# $0.30/$1.20 rate as the first-party API. Verified 2026-07-02.
[models."MiniMaxAI/MiniMax-M3"]
name = "MiniMax M3 (Together)"
provider = "together"
context_window = 524288
capabilities = ["tools", "vision", "video", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.30, output_per_mtok = 1.20, cache_read_per_mtok = 0.06 }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "tool_use", "reasoning", "long_context", "vision"]

# Kimi K2.7 Code on Together — int4-native weights at the first-party
# list rate. Verified 2026-07-02.
[models."moonshotai/Kimi-K2.7-Code"]
name = "Kimi K2.7 Code (Together)"
provider = "together"
context_window = 262144
capabilities = ["tools", "vision", "video", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.95, output_per_mtok = 4.00, cache_read_per_mtok = 0.19 }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "long_context", "tool_use", "reasoning"]
[models."MiniMaxAI/MiniMax-M2.7"]
name = "MiniMax M2.7 (Together)"
provider = "together"
context_window = 196608
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.30, output_per_mtok = 1.20, cache_read_per_mtok = 0.06 }
# Inline metadata matches the minimax-m2.7 equivalence group and keeps this row
# independent of fragment-order defaults.
tier = "frontier"
open_weight = true
strengths = ["speed", "coding", "agentic", "tool_use", "reasoning", "long_context"]

# Reliable Together serverless sample routes (harn#3645 §4). The prior
# representative serverless rows were unusable as one-click samples:
# `Qwen/Qwen3-Coder-Next-FP8` returns `model_not_available` on serverless
# chat-completions (needs a dedicated endpoint; carried as
# `availability = "dedicated"` in 20-open-weight-openrouter.toml), and the
# Together `google/gemma-4-31B-it` route is a reasoning model with an
# empty-content footgun. These two non-reasoning Turbo routes are live on
# Together's serverless `/v1/chat/completions` with native tool calls and
# serve as dependable cheap executor samples. Verified live ids on
# together.ai/models (Qwen2.5 7B Instruct Turbo, Llama 3.3 70B Instruct
# Turbo).
[models."Qwen/Qwen2.5-7B-Instruct-Turbo"]
name = "Qwen2.5 7B Instruct Turbo (Together)"
provider = "together"
context_window = 32768
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.30, output_per_mtok = 0.30 }
tier = "small"
open_weight = true
strengths = ["speed", "cheap", "tool_use"]
[models."meta-llama/Llama-3.3-70B-Instruct-Turbo"]
name = "Llama 3.3 70B Instruct Turbo (Together)"
provider = "together"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.88, output_per_mtok = 0.88 }
logical_model = "llama-3.3-70b-instruct"
equivalence_group = "llama-3.3-70b-instruct"
served_variant = "together"
api_dialect = "openai_chat"
tier = "mid"
open_weight = true
strengths = ["tool_use", "cheap"]