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
122
123
124

# llama.cpp — Unsloth Dynamic 2.0 GGUF served by llama-server.
[models."qwen3.6-35b-a3b-ud-q4-k-xl"]
name = "Qwen3.6 35B (Unsloth Q4_K_XL, llama.cpp)"
provider = "llamacpp"
context_window = 262144
runtime_context_window = 262144
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]
tier = "mid"
open_weight = true
strengths = ["coding"]
[models."qwen3.6-35b-a3b-ud-q4-k-xl".local_memory]
measured_resident_gib = 19.5
measured_context_window = 8192
measured_cache_type = "q8_0"
base_resident_gib = 19.0
kv_cache_gib_per_1k_ctx = 0.0135
default_cache_type = "q8_0"
safety_margin_gib = 4.0
max_recommended_context = 262144
cache_type_multipliers = { q8_0 = 1.0, f16 = 2.0, q4_0 = 0.5, q4_1 = 0.5, q5_0 = 0.625, q5_1 = 0.625 }
last_verified = "2026-06-28"
notes = "Empirical llama.cpp b9821 / CUDA 12.8 / RTX 5090 measurements with q8_0 KV showed SWA-capped KV growth of about 0.87 GiB per 65k tokens and full 262144 ctx fitting with headroom. Treat as a sizing estimate, not an allocator guarantee."
[models."qwen3.6-35b-a3b-ud-q5-k-xl"]
name = "Qwen3.6 35B (Unsloth Q5_K_XL, llama.cpp)"
provider = "llamacpp"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]
tier = "mid"
open_weight = true
strengths = ["coding"]
[models."qwen3.6-35b-a3b"]
name = "Qwen3.6 35B (llama.cpp)"
provider = "llamacpp"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]
tier = "mid"
open_weight = true
strengths = ["coding"]

# Apple Silicon MLX. Burin #2717 switched the local MLX route from the
# never-downloaded dense vision model `unsloth/Qwen3.6-27B-UD-MLX-4bit` to the
# coding-tuned Qwen3.6-35B-A3B MoE served via `mlx_lm.server` (text MoE, no
# vision). Shares the `qwen3.6-35b-a3b` logical_model / equivalence_group with
# the llama.cpp GGUF route so eval aggregation treats the two runtimes as the
# same model and compares them directly. Keep strengths conservative until
# `mlx_lm.server` has its own tool/agentic probe evidence; equivalence groups
# use the least-decorated host baseline for escalation decisions.
[models."unsloth/Qwen3.6-35B-A3B-UD-MLX-4bit"]
name = "Qwen3.6 35B-A3B (MLX 4-bit)"
provider = "mlx"
logical_model = "qwen3.6-35b-a3b"
equivalence_group = "qwen3.6-35b-a3b"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]
tier = "mid"
open_weight = true
strengths = ["coding"]

[models."unsloth/Qwen3.6-35B-A3B-UD-MLX-8bit"]
name = "Qwen3.6 35B-A3B (MLX 8-bit)"
provider = "mlx"
logical_model = "qwen3.6-35b-a3b"
equivalence_group = "qwen3.6-35b-a3b"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]
tier = "mid"
open_weight = true
strengths = ["coding"]

# Local OpenAI-compatible servers (vLLM / bring-your-own).
[models."gemma-4-e2b-it"]
name = "Gemma 4 E2B (local)"
provider = "local"
context_window = 131072
stream_timeout = 300.0
capabilities = ["streaming", "tools", "vision", "thinking", "structured_output"]
tier = "small"
open_weight = true
strengths = ["cheap", "speed"]
[models."gemma-4-e4b-it"]
name = "Gemma 4 E4B (local)"
provider = "local"
context_window = 131072
stream_timeout = 300.0
capabilities = ["streaming", "tools", "vision", "thinking", "structured_output"]
tier = "small"
open_weight = true
strengths = ["cheap"]
[models."gemma-4-26b-a4b-it"]
name = "Gemma 4 26B MoE (local)"
provider = "local"
context_window = 131072
stream_timeout = 600.0
capabilities = ["streaming", "tools", "vision", "thinking", "structured_output"]
tier = "mid"
open_weight = true
strengths = ["coding"]
[models."gemma-4-31b-it"]
name = "Gemma 4 31B (local)"
provider = "local"
context_window = 131072
stream_timeout = 600.0
capabilities = ["streaming", "tools", "vision", "thinking", "structured_output"]
tier = "frontier"
open_weight = true
strengths = ["coding", "long_context"]
[models."gemma-4-12b-it"]
name = "Gemma 4 12B (local)"
provider = "local"
context_window = 131072
stream_timeout = 300.0
capabilities = ["streaming", "tools", "vision", "thinking", "structured_output"]
tier = "mid"
open_weight = true
strengths = ["cheap", "speed"]