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

# Open-weight executor candidates (<$2/Mtok with function calling). Use
# these via OpenRouter or Fireworks for fast secondary-model dispatch.
# Pricing snapshot 2026-05 from OpenRouter / Artificial Analysis.
[models."qwen/qwen3-coder"]
name = "Qwen3 Coder 480B A35B"
provider = "openrouter"
context_window = 262144
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.22, output_per_mtok = 1.80 }
availability = "serverless"
tier = "frontier"
open_weight = true
strengths = ["coding", "long_context", "agentic", "tool_use"]
benchmarks = { swe_bench_verified = 67.0 }

# Qwen3-Coder-Next — 80B/3B MoE non-thinking coder.
[models."qwen/qwen3-coder-next"]
name = "Qwen3 Coder Next"
provider = "openrouter"
context_window = 262144
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.11, output_per_mtok = 0.80, cache_read_per_mtok = 0.07 }
availability = "serverless"
architecture = { parameter_count_b = 80.0, active_parameter_count_b = 3.0, moe = true, license = "Apache-2.0", source_url = "https://huggingface.co/Qwen/Qwen3-Coder-Next", last_verified = "2026-07-02" }
benchmarks = { swe_bench_verified = 70.6, swe_bench_pro = 44.3 }
tier = "mid"
open_weight = true
strengths = ["coding", "cheap", "long_context", "agentic", "tool_use"]

# Qwen3.5-397B-A17B — the open-tier Qwen flagship (native VLM MoE).
# OpenRouter serverless rate verified 2026-07-02.
[models."qwen/qwen3.5-397b-a17b"]
name = "Qwen3.5 397B A17B"
provider = "openrouter"
context_window = 262144
capabilities = ["tools", "vision", "streaming", "thinking"]
pricing = { input_per_mtok = 0.385, output_per_mtok = 2.45 }
availability = "serverless"
architecture = { parameter_count_b = 397.0, active_parameter_count_b = 17.0, moe = true, license = "Apache-2.0", source_url = "https://huggingface.co/Qwen/Qwen3.5-397B-A17B", last_verified = "2026-07-02" }
benchmarks = { swe_bench_verified = 76.4 }
tier = "frontier"
open_weight = true
strengths = ["coding", "reasoning", "tool_use", "long_context", "vision"]

# Together lists Qwen3-Coder-Next-FP8 in GET /v1/models alongside its
# serverless catalog, but normal chat-completion calls fail with
# `model_not_available` and instruct the caller to create a dedicated
# endpoint. Carry the catalog row so price/route metadata is preserved,
# but mark `availability = "dedicated"` so hosts don't surface it as a
# one-click serverless option.
[models."Qwen/Qwen3-Coder-Next-FP8"]
name = "Qwen3 Coder Next FP8 (Together, dedicated)"
provider = "together"
context_window = 262144
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.18, output_per_mtok = 0.18 }
availability = "dedicated"
tier = "frontier"
open_weight = true
strengths = ["coding", "long_context", "agentic"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "deepseek", "kimi"]
[models."deepseek/deepseek-v3.2"]
name = "DeepSeek V3.2"
provider = "openrouter"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.2288, output_per_mtok = 0.3432 }
tier = "mid"
open_weight = true
strengths = ["coding", "tool_use", "cheap"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "kimi"]
[models."moonshotai/kimi-k2.6"]
name = "Kimi K2.6"
provider = "openrouter"
context_window = 262144
capabilities = ["tools", "vision", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.68, output_per_mtok = 3.41, cache_read_per_mtok = 0.34 }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "long_context", "tool_use", "reasoning", "vision"]
benchmarks = { swe_bench_pro = 58.6, humanitys_last_exam_with_tools = 54.0 }
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]
[models."moonshotai/kimi-k2.7-code"]
name = "Kimi K2.7 Code"
provider = "openrouter"
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", "vision"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]
[models."openai/gpt-oss-120b"]
name = "GPT-OSS 120B"
provider = "openrouter"
context_window = 131072
logical_model = "openai-gpt-oss-120b"
equivalence_group = "openai-gpt-oss-120b"
served_variant = "openrouter"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.15, output_per_mtok = 0.60 }
architecture = { parameter_count_b = 117.0, active_parameter_count_b = 5.1, moe = true, license = "Apache-2.0", source_url = "https://developers.openai.com/api/docs/models/gpt-oss-120b", last_verified = "2026-06-05" }
# Explicit metadata keeps this row self-described across fragment boundaries.
# GPT-OSS 120B is Apache-2.0 (open weight) and text-only; it is Burin's cheap
# coding executor.
# `tier`/`open_weight`/`strengths` match the rest of the openai-gpt-oss-120b
# equivalence group (the validator requires one tier per logical model — the
# conservative shared baseline is `frontier`).
tier = "frontier"
open_weight = true
strengths = ["cheap", "coding", "tool_use", "reasoning"]