{"id":475496,"date":"2026-04-11T10:33:15","date_gmt":"2026-04-11T10:33:15","guid":{"rendered":"https:\/\/savepearlharbor.com\/?p=475496"},"modified":"-0001-11-30T00:00:00","modified_gmt":"-0001-11-29T21:00:00","slug":"","status":"publish","type":"post","link":"https:\/\/savepearlharbor.com\/?p=475496","title":{"rendered":"\u041b\u043e\u043a\u0430\u043b\u044c\u043d\u044b\u0439 \u0437\u0430\u043f\u0443\u0441\u043a GLM-5.1"},"content":{"rendered":"
\n

\u041f\u0435\u0440\u0435\u0432\u043e\u0434 \u043f\u043e\u0434\u0433\u043e\u0442\u043e\u0432\u0438\u043b \u0430\u0432\u0442\u043e\u0440 \u043a\u0430\u043d\u0430\u043b\u0430<\/em>\u00a0\u0414\u0440\u0443\u0433 \u041e\u043f\u0435\u043d\u0441\u0443\u0440\u0441\u0430<\/em><\/a>, \u043f\u0440\u0438\u044f\u0442\u043d\u043e\u0433\u043e \u043f\u0440\u043e\u0447\u0442\u0435\u043d\u0438\u044f, \u0437\u0430\u0440\u0430\u043d\u0435\u0435 \u0431\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044e \u0437\u0430 \u043f\u043e\u0434\u043f\u0438\u0441\u043a\u0443<\/em> <\/p>\n

GLM-5.1 \u2014 \u044d\u0442\u043e \u043d\u043e\u0432\u0430\u044f \u043e\u0442\u043a\u0440\u044b\u0442\u0430\u044f \u043c\u043e\u0434\u0435\u043b\u044c \u043e\u0442 Z.ai. \u041e\u043d\u0430 \u0438\u043c\u0435\u0435\u0442 744 \u043c\u043b\u0440\u0434 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 (40 \u043c\u043b\u0440\u0434 \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445) \u0438 \u043a\u043e\u043d\u0442\u0435\u043a\u0441\u0442\u043d\u043e\u0435 \u043e\u043a\u043d\u043e 200K. \u041f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 GLM-5 \u0432 \u043d\u0435\u0439 \u0443\u043b\u0443\u0447\u0448\u0435\u043d\u044b \u043d\u0430\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u043a\u043e\u0434\u0430, \u0440\u0430\u0431\u043e\u0442\u0430 \u0441 \u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442\u0430\u043c\u0438 \u0438 \u043b\u043e\u0433\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0437\u0430\u0434\u0430\u0447\u0438. <\/p>\n

<\/figure>\n
<\/figure>\n

\u0414\u043b\u044f \u0445\u0440\u0430\u043d\u0435\u043d\u0438\u044f \u043f\u043e\u043b\u043d\u043e\u0439 \u043c\u043e\u0434\u0435\u043b\u0438 \u043d\u0443\u0436\u043d\u043e 1,65 \u0422\u0411. \u041a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u043d\u044b\u0435 \u0432\u0435\u0440\u0441\u0438\u0438 GGUF \u043e\u0442 Unsloth \u0437\u0430\u043d\u0438\u043c\u0430\u044e\u0442 \u043c\u0435\u043d\u044c\u0448\u0435 \u043c\u0435\u0441\u0442\u0430: 2-\u0431\u0438\u0442\u043d\u0430\u044f \u0432\u0435\u0440\u0441\u0438\u044f \u0442\u0440\u0435\u0431\u0443\u0435\u0442 220 \u0413\u0411, 1-\u0431\u0438\u0442\u043d\u0430\u044f 200 \u0413\u0411. <\/p>\n

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

Benchmark<\/p>\n<\/th>\n

\n

GLM-5.1<\/p>\n<\/th>\n

\n

GLM-5<\/p>\n<\/th>\n

\n

Qwen3.6+<\/p>\n<\/th>\n

\n

Minimax M2.7<\/p>\n<\/th>\n

\n

DeepSeek V3.2<\/p>\n<\/th>\n

\n

Kimi K2.5<\/p>\n<\/th>\n

\n

Claude Opus 4.6<\/p>\n<\/th>\n

\n

Gemini 3.1 Pro<\/p>\n<\/th>\n

\n

GPT-5.4<\/p>\n<\/th>\n<\/tr>\n

\n

HLE<\/strong><\/p>\n<\/td>\n

\n

31.0<\/p>\n<\/td>\n

\n

30.5<\/p>\n<\/td>\n

\n

28.8<\/p>\n<\/td>\n

\n

28.0<\/p>\n<\/td>\n

\n

25.1<\/p>\n<\/td>\n

\n

31.5<\/p>\n<\/td>\n

\n

36.7<\/p>\n<\/td>\n

\n

45.0<\/p>\n<\/td>\n

\n

39.8<\/p>\n<\/td>\n<\/tr>\n

\n

HLE (Tools)<\/strong><\/p>\n<\/td>\n

\n

52.3<\/p>\n<\/td>\n

\n

50.4<\/p>\n<\/td>\n

\n

50.6<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

40.8<\/p>\n<\/td>\n

\n

51.8<\/p>\n<\/td>\n

\n

53.1*<\/p>\n<\/td>\n

\n

51.4*<\/p>\n<\/td>\n

\n

52.1*<\/p>\n<\/td>\n<\/tr>\n

\n

AIME 2026<\/strong><\/p>\n<\/td>\n

\n

95.3<\/p>\n<\/td>\n

\n

95.4<\/p>\n<\/td>\n

\n

95.1<\/p>\n<\/td>\n

\n

89.8<\/p>\n<\/td>\n

\n

95.1<\/p>\n<\/td>\n

\n

94.5<\/p>\n<\/td>\n

\n

95.6<\/p>\n<\/td>\n

\n

98.2<\/p>\n<\/td>\n

\n

98.7<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

HMMT Nov 2025<\/strong><\/p>\n<\/td>\n

\n

94.0<\/p>\n<\/td>\n

\n

96.9<\/strong><\/p>\n<\/td>\n

\n

94.6<\/p>\n<\/td>\n

\n

81.0<\/p>\n<\/td>\n

\n

90.2<\/p>\n<\/td>\n

\n

91.1<\/p>\n<\/td>\n

\n

96.3<\/p>\n<\/td>\n

\n

94.8<\/p>\n<\/td>\n

\n

95.8<\/p>\n<\/td>\n<\/tr>\n

\n

HMMT Feb 2026<\/strong><\/p>\n<\/td>\n

\n

82.6<\/p>\n<\/td>\n

\n

82.8<\/p>\n<\/td>\n

\n

87.8<\/p>\n<\/td>\n

\n

72.7<\/p>\n<\/td>\n

\n

79.9<\/p>\n<\/td>\n

\n

81.3<\/p>\n<\/td>\n

\n

84.3<\/p>\n<\/td>\n

\n

87.3<\/p>\n<\/td>\n

\n

91.8<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

IMOAnswerBench<\/strong><\/p>\n<\/td>\n

\n

83.8<\/p>\n<\/td>\n

\n

82.5<\/p>\n<\/td>\n

\n

83.8<\/p>\n<\/td>\n

\n

66.3<\/p>\n<\/td>\n

\n

78.3<\/p>\n<\/td>\n

\n

81.8<\/p>\n<\/td>\n

\n

75.3<\/p>\n<\/td>\n

\n

81.0<\/p>\n<\/td>\n

\n

91.4<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

GPQA-Diamond<\/strong><\/p>\n<\/td>\n

\n

86.2<\/p>\n<\/td>\n

\n

86.0<\/p>\n<\/td>\n

\n

90.4<\/p>\n<\/td>\n

\n

87.0<\/p>\n<\/td>\n

\n

82.4<\/p>\n<\/td>\n

\n

87.6<\/p>\n<\/td>\n

\n

91.3<\/p>\n<\/td>\n

\n

94.3<\/strong><\/p>\n<\/td>\n

\n

92.0<\/p>\n<\/td>\n<\/tr>\n

\n

SWE-Bench Pro<\/strong><\/p>\n<\/td>\n

\n

58.4<\/strong><\/p>\n<\/td>\n

\n

55.1<\/p>\n<\/td>\n

\n

56.6<\/p>\n<\/td>\n

\n

56.2<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

53.8<\/p>\n<\/td>\n

\n

57.3<\/p>\n<\/td>\n

\n

54.2<\/p>\n<\/td>\n

\n

57.7<\/p>\n<\/td>\n<\/tr>\n

\n

NL2Repo<\/strong><\/p>\n<\/td>\n

\n

42.7<\/p>\n<\/td>\n

\n

35.9<\/p>\n<\/td>\n

\n

37.9<\/p>\n<\/td>\n

\n

39.8<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

32.0<\/p>\n<\/td>\n

\n

49.8<\/strong><\/p>\n<\/td>\n

\n

33.4<\/p>\n<\/td>\n

\n

41.3<\/p>\n<\/td>\n<\/tr>\n

\n

TerminalBench 2<\/strong><\/p>\n<\/td>\n

\n

63.5<\/p>\n<\/td>\n

\n

56.2<\/p>\n<\/td>\n

\n

61.6<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

39.3<\/p>\n<\/td>\n

\n

50.8<\/p>\n<\/td>\n

\n

65.4<\/p>\n<\/td>\n

\n

68.5<\/strong><\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n<\/tr>\n

\n

TerminalBench (Best)<\/strong><\/p>\n<\/td>\n

\n

66.5 (Claude Code)<\/p>\n<\/td>\n

\n

56.2
(Claude Code)<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

57.0
(Claude Code)<\/p>\n<\/td>\n

\n

46.4
(Claude Code)<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

75.1<\/strong>
(Codex)<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

CyberGym<\/strong><\/p>\n<\/td>\n

\n

68.7<\/strong><\/p>\n<\/td>\n

\n

48.3<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

17.3<\/p>\n<\/td>\n

\n

41.3<\/p>\n<\/td>\n

\n

66.6<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n<\/tr>\n

\n

BrowseComp<\/strong><\/p>\n<\/td>\n

\n

68.0<\/strong><\/p>\n<\/td>\n

\n

62.0<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

51.4<\/p>\n<\/td>\n

\n

60.6<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n<\/tr>\n

\n

BrowseComp (CM)<\/strong><\/p>\n<\/td>\n

\n

79.3<\/p>\n<\/td>\n

\n

75.9<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

67.6<\/p>\n<\/td>\n

\n

74.9<\/p>\n<\/td>\n

\n

84.0<\/p>\n<\/td>\n

\n

85.9<\/strong><\/p>\n<\/td>\n

\n

82.7<\/p>\n<\/td>\n<\/tr>\n

\n

\u03c4\u00b3-Bench<\/strong><\/p>\n<\/td>\n

\n

70.6<\/p>\n<\/td>\n

\n

69.2<\/p>\n<\/td>\n

\n

70.7<\/p>\n<\/td>\n

\n

67.6<\/p>\n<\/td>\n

\n

69.2<\/p>\n<\/td>\n

\n

66.0<\/p>\n<\/td>\n

\n

72.4<\/p>\n<\/td>\n

\n

67.1<\/p>\n<\/td>\n

\n

72.9<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

MCP-Atlas<\/strong><\/p>\n<\/td>\n

\n

71.8<\/p>\n<\/td>\n

\n

69.2<\/p>\n<\/td>\n

\n

74.1<\/strong><\/p>\n<\/td>\n

\n

48.8<\/p>\n<\/td>\n

\n

62.2<\/p>\n<\/td>\n

\n

63.8<\/p>\n<\/td>\n

\n

73.8<\/p>\n<\/td>\n

\n

69.2<\/p>\n<\/td>\n

\n

67.2<\/p>\n<\/td>\n<\/tr>\n

\n

Tool-Decathlon<\/strong><\/p>\n<\/td>\n

\n

40.7<\/p>\n<\/td>\n

\n

38.0<\/p>\n<\/td>\n

\n

39.8<\/p>\n<\/td>\n

\n

46.3<\/p>\n<\/td>\n

\n

35.2<\/p>\n<\/td>\n

\n

27.8<\/p>\n<\/td>\n

\n

47.2<\/p>\n<\/td>\n

\n

48.8<\/p>\n<\/td>\n

\n

54.6<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

Vending Bench 2 ($)<\/strong><\/p>\n<\/td>\n

\n

$5634<\/p>\n<\/td>\n

\n

$4432<\/p>\n<\/td>\n

\n

$5115<\/p>\n<\/td>\n

\n

\u2014<\/p>\n<\/td>\n

\n

$1034<\/p>\n<\/td>\n

\n

$1198<\/p>\n<\/td>\n

\n

$8018<\/strong><\/p>\n<\/td>\n

\n

$911<\/p>\n<\/td>\n

\n

$6144<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u041d\u043e \u043d\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0439\u0442\u0435 CUDA 13.2 \u0434\u043b\u044f GGUF, \u044d\u0442\u043e \u043f\u043e\u0440\u0442\u0438\u0442 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u043e\u0442\u0432\u0435\u0442\u043e\u0432. <\/p>\n\n

\u0422\u0440\u0435\u0431\u043e\u0432\u0430\u043d\u0438\u044f \u043a \u043e\u0431\u043e\u0440\u0443\u0434\u043e\u0432\u0430\u043d\u0438\u044e<\/h3>\n

\u0412\u0435\u0440\u0441\u0438\u044f UD-IQ2_M \u043f\u043e\u0434\u0445\u043e\u0434\u0438\u0442 \u0434\u043b\u044f Mac \u0441 256 \u0413\u0411 \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u0438. \u0422\u0430\u043a\u0436\u0435 \u043e\u043d\u0430 \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043d\u0430 \u041f\u041a \u0441 \u043e\u0434\u043d\u043e\u0439 \u0432\u0438\u0434\u0435\u043e\u043a\u0430\u0440\u0442\u043e\u0439 \u043d\u0430 24 \u0413\u0411 \u0438 256 \u0413\u0411 \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u0438 \u0447\u0435\u0440\u0435\u0437 MoE offloading. \u0414\u043b\u044f 1-\u0431\u0438\u0442\u043d\u043e\u0439 \u0432\u0435\u0440\u0441\u0438\u0438 \u043d\u0443\u0436\u043d\u043e 220 \u0413\u0411 \u043f\u0430\u043c\u044f\u0442\u0438, \u0434\u043b\u044f 8-\u0431\u0438\u0442\u043d\u043e\u0439 805 \u0413\u0411. <\/p>\n

\n
\n\n\n\n\n\n\n
\n

\u041f\u0430\u0440\u0430\u043c\u0435\u0442\u0440<\/p>\n<\/th>\n

\n

Default Settings (Most Tasks)<\/p>\n<\/th>\n

\n

Terminal Bench<\/p>\n<\/th>\n<\/tr>\n

\n

temperature<\/strong><\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.7<\/p>\n<\/td>\n<\/tr>\n

\n

top_p<\/strong><\/p>\n<\/td>\n

\n

0.95<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n<\/tr>\n

\n

max new tokens<\/strong><\/p>\n<\/td>\n

\n

131072<\/p>\n<\/td>\n

\n

16384<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u0412 GLM-5.1 \u043f\u043e \u0443\u043c\u043e\u043b\u0447\u0430\u043d\u0438\u044e \u0432\u043a\u043b\u044e\u0447\u0435\u043d \u0440\u0435\u0436\u0438\u043c \u0440\u0430\u0441\u0441\u0443\u0436\u0434\u0435\u043d\u0438\u044f. \u0414\u043b\u044f \u0435\u0433\u043e \u043e\u0442\u043a\u043b\u044e\u0447\u0435\u043d\u0438\u044f \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0439\u0442\u0435<\/p>\n

--chat-template-kwargs '{\"enable_thinking\":false}'<\/code>
<\/a><\/div><\/pre>\n
\u0418\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 Unsloth Studio<\/h3>\n
  \u0423\u0441\u0442\u0430\u043d\u043e\u0432\u043a\u0430 \u0434\u043b\u044f MacOS, Linux, WSL<\/strong><\/p>\n
curl -fsSL https:\/\/unsloth.ai\/install.sh | sh<\/code>
<\/a><\/div><\/pre>\n
\u0423\u0441\u0442\u0430\u043d\u043e\u0432\u043a\u0430 \u0434\u043b\u044f Windows PowerShell <\/strong><\/p>\n
irm https:\/\/unsloth.ai\/install.ps1 | iex<\/code>
<\/a><\/div><\/pre>\n
\u0417\u0430\u043f\u0443\u0441\u043a<\/strong><\/p>\n
unsloth studio -H 0.0.0.0 -p 8888<\/code>
<\/a><\/div><\/pre>\n
\u041f\u043e\u0441\u043b\u0435 \u0437\u0430\u043f\u0443\u0441\u043a\u0430 \u043e\u0442\u043a\u0440\u043e\u0439\u0442\u0435 \u00a0http:\/\/localhost:8888<\/code><\/a>   \u0412 \u043f\u043e\u0438\u0441\u043a\u0435 \u043d\u0430\u0439\u0434\u0438\u0442\u0435 GLM-5.1. \u0420\u0435\u043a\u043e\u043c\u0435\u043d\u0434\u0443\u0435\u0442\u0441\u044f \u0432\u044b\u0431\u0438\u0440\u0430\u0442\u044c \u0432\u0435\u0440\u0441\u0438\u044e UD-Q2_K_XL. \u0415\u0441\u043b\u0438 VRAM \u043d\u0435 \u0445\u0432\u0430\u0442\u0438\u0442, \u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u0437\u0430\u0434\u0435\u0439\u0441\u0442\u0432\u0443\u0435\u0442 \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u0443\u044e \u043f\u0430\u043c\u044f\u0442\u044c.<\/p>\n
<\/figure>\n\n
\u0418\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 llama.cpp<\/h3>\n
  \u0421\u043e\u0431\u0435\u0440\u0438\u0442\u0435 llama.cpp \u0438\u0437 \u0440\u0435\u043f\u043e\u0437\u0438\u0442\u043e\u0440\u0438\u044f GitHub. \u0414\u043b\u044f \u0441\u0438\u0441\u0442\u0435\u043c \u0431\u0435\u0437 GPU \u0438\u043b\u0438 \u043d\u0430 Mac \u0441 Metal \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0439\u0442\u0435\u00a0-DGGML_CUDA=OFF<\/code>.  <\/p>\n
  \u0414\u043b\u044f \u0441\u0431\u043e\u0440\u043a\u0438 <\/strong><\/p>\n
apt-get updateapt-get install pciutils build-essential cmake curl libcurl4-openssl-dev -ygit clone https:\/\/github.com\/ggml-org\/llama.cppcmake llama.cpp -B llama.cpp\/build \\    -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ONcmake --build llama.cpp\/build --config Release -j --clean-first --target llama-cli llama-mtmd-cli llama-server llama-gguf-splitcp llama.cpp\/build\/bin\/llama-* llama.cpp<\/code>
<\/a><\/div><\/pre>\n
  \u0417\u0430\u043f\u0443\u0441\u043a \u043e\u0431\u044b\u0447\u043d\u043e\u0439 \u0432\u0435\u0440\u0441\u0438\u0438 <\/strong><\/p>\n
export LLAMA_CACHE=\"unsloth\/GLM-5.1-GGUF\".\/llama.cpp\/llama-cli \\    -hf unsloth\/GLM-5.1-GGUF:UD-IQ2_M \\    --ctx-size 16384 \\    --temp 0.7 \\    --top-p 1.0<\/code>
<\/a><\/div><\/pre>\n
  \u0417\u0430\u043f\u0443\u0441\u043a \u0434\u043b\u044f \u0432\u044b\u0437\u043e\u0432\u0430 \u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442\u043e\u0432<\/strong><\/p>\n
export LLAMA_CACHE=\"unsloth\/GLM-5.1-GGUF\".\/llama.cpp\/llama-cli \\    -hf unsloth\/GLM-5.1-GGUF:UD-IQ2_M \\    --ctx-size 16384 \\    --temp 1.0 \\    --top-p 0.95<\/code>
<\/a><\/div><\/pre>\n
\u0420\u0430\u0431\u043e\u0442\u0430 \u0447\u0435\u0440\u0435\u0437 API<\/h3>\n
  \u0414\u043b\u044f \u0437\u0430\u043f\u0443\u0441\u043a\u0430 \u0441\u0435\u0440\u0432\u0435\u0440\u0430 <\/strong><\/p>\n
.\/llama.cpp\/llama-server \\    --model unsloth\/GLM-5.1-GGUF\/UD-IQ2_M\/GLM-5.1-UD-IQ2_M-00001-of-00006.gguf \\    --alias \"unsloth\/GLM-5.1\" \\    --prio 3 \\    --temp 1.0 \\    --top-p 0.95 \\    --ctx-size 16384 \\    --port 8001<\/code>
<\/a><\/div><\/pre>\n
  \u041f\u0440\u0438\u043c\u0435\u0440 \u043a\u043e\u0434\u0430 \u043d\u0430 Python \u0434\u043b\u044f \u043e\u0431\u0440\u0430\u0449\u0435\u043d\u0438\u044f \u043a \u0441\u0435\u0440\u0432\u0435\u0440\u0443<\/strong><\/p>\n
from openai import OpenAIopenai_client = OpenAI(    base_url = \"http:\/\/127.0.0.1:8001\/v1\",    api_key = \"sk-no-key-required\",)completion = openai_client.chat.completions.create(    model = \"unsloth\/GLM-5.1\",    messages = [{\"role\": \"user\", \"content\": \"Create a Snake game.\"},],)print(completion.choices[0].message.content)<\/code>
<\/a><\/div><\/pre>\n
\u0412\u044b\u0437\u043e\u0432 \u0444\u0443\u043d\u043a\u0446\u0438\u0439<\/h3>\n
GLM-5.1 \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442 \u0432\u044b\u0437\u043e\u0432 \u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442\u043e\u0432. \u0414\u043b\u044f \u044d\u0442\u043e\u0433\u043e \u043d\u0443\u0436\u043d\u043e \u043e\u043f\u0440\u0435\u0434\u0435\u043b\u0438\u0442\u044c \u0444\u0443\u043d\u043a\u0446\u0438\u0438 \u0432 Python \u0438 \u043f\u0435\u0440\u0435\u0434\u0430\u0442\u044c \u0438\u0445 \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u0432 \u043c\u043e\u0434\u0435\u043b\u044c. <\/p>\n
def add_number(a: float | str, b: float | str) -> float:    return float(a) + float(b)def terminal(command: str) -> str:    if \"rm\" in command or \"sudo\" in command or \"dd\" in command or \"chmod\" in command:        return \"Dangerous command\"    return str(subprocess.run(command, capture_output = True, text = True, shell = True).stdout)<\/code>
<\/a><\/div><\/pre>\n
\u041c\u043e\u0434\u0435\u043b\u044c \u043c\u043e\u0436\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u044d\u0442\u0438 \u0444\u0443\u043d\u043a\u0446\u0438\u0438 \u0434\u043b\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0440\u0430\u0441\u0447\u0435\u0442\u043e\u0432 \u0438\u043b\u0438 \u043a\u043e\u043c\u0430\u043d\u0434 \u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u0435. \u041f\u0440\u0438 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0438\u00a0llama-server<\/code>\u00a0\u0430\u0432\u0442\u043e\u043c\u0430\u0442\u0438\u0447\u0435\u0441\u043a\u0430\u044f \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0430 \u0432\u044b\u0437\u043e\u0432\u043e\u0432 \u043d\u0430\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0435\u0442\u0441\u044f \u0447\u0435\u0440\u0435\u0437 \u0446\u0438\u043a\u043b, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u043f\u0440\u043e\u0432\u0435\u0440\u044f\u0435\u0442 \u043d\u0430\u043b\u0438\u0447\u0438\u0435\u00a0tool_calls<\/code>\u00a0\u0432 \u043e\u0442\u0432\u0435\u0442\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 \u0438 \u0432\u043e\u0437\u0432\u0440\u0430\u0449\u0430\u0435\u0442 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0444\u0443\u043d\u043a\u0446\u0438\u0438.<\/p>\n<\/div>\n
\u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438 https:\/\/habr.com\/ru\/articles\/1022242\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
\u041f\u0435\u0440\u0435\u0432\u043e\u0434 \u043f\u043e\u0434\u0433\u043e\u0442\u043e\u0432\u0438\u043b \u0430\u0432\u0442\u043e\u0440 \u043a\u0430\u043d\u0430\u043b\u0430\u00a0\u0414\u0440\u0443\u0433 \u041e\u043f\u0435\u043d\u0441\u0443\u0440\u0441\u0430, \u043f\u0440\u0438\u044f\u0442\u043d\u043e\u0433\u043e \u043f\u0440\u043e\u0447\u0442\u0435\u043d\u0438\u044f, \u0437\u0430\u0440\u0430\u043d\u0435\u0435 \u0431\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044e \u0437\u0430 \u043f\u043e\u0434\u043f\u0438\u0441\u043a\u0443  GLM-5.1 \u2014 \u044d\u0442\u043e \u043d\u043e\u0432\u0430\u044f \u043e\u0442\u043a\u0440\u044b\u0442\u0430\u044f \u043c\u043e\u0434\u0435\u043b\u044c \u043e\u0442 Z.ai. \u041e\u043d\u0430 \u0438\u043c\u0435\u0435\u0442 744 \u043c\u043b\u0440\u0434 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 (40 \u043c\u043b\u0440\u0434 \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445) \u0438 \u043a\u043e\u043d\u0442\u0435\u043a\u0441\u0442\u043d\u043e\u0435 \u043e\u043a\u043d\u043e 200K. \u041f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 GLM-5 \u0432 \u043d\u0435\u0439 \u0443\u043b\u0443\u0447\u0448\u0435\u043d\u044b \u043d\u0430\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u043a\u043e\u0434\u0430, \u0440\u0430\u0431\u043e\u0442\u0430 \u0441 \u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442\u0430\u043c\u0438 \u0438 \u043b\u043e\u0433\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0437\u0430\u0434\u0430\u0447\u0438.  \u0414\u043b\u044f \u0445\u0440\u0430\u043d\u0435\u043d\u0438\u044f \u043f\u043e\u043b\u043d\u043e\u0439 \u043c\u043e\u0434\u0435\u043b\u0438 \u043d\u0443\u0436\u043d\u043e 1,65 \u0422\u0411. \u041a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u043d\u044b\u0435 \u0432\u0435\u0440\u0441\u0438\u0438 GGUF \u043e\u0442 Unsloth \u0437\u0430\u043d\u0438\u043c\u0430\u044e\u0442 \u043c\u0435\u043d\u044c\u0448\u0435 \u043c\u0435\u0441\u0442\u0430: 2-\u0431\u0438\u0442\u043d\u0430\u044f \u0432\u0435\u0440\u0441\u0438\u044f \u0442\u0440\u0435\u0431\u0443\u0435\u0442 220 \u0413\u0411, 1-\u0431\u0438\u0442\u043d\u0430\u044f  200 \u0413\u0411.  BenchmarkGLM-5.1GLM-5Qwen3.6+Minimax M2.7DeepSeek V3.2Kimi K2.5Claude Opus 4.6Gemini 3.1 ProGPT-5.4HLE31.030.528.828.025.131.536.745.039.8HLE (Tools)52.350.450.6\u201440.851.853.1*51.4*52.1*AIME 202695.395.495.189.895.194.595.698.298.7HMMT Nov 202594.096.994.681.090.291.196.394.895.8HMMT Feb 202682.682.887.872.779.981.384.387.391.8IMOAnswerBench83.882.583.866.378.381.875.381.091.4GPQA-Diamond86.286.090.487.082.487.691.394.392.0SWE-Bench Pro58.455.156.656.2\u201453.857.354.257.7NL2Repo42.735.937.939.8\u201432.049.833.441.3TerminalBench 263.556.261.6\u201439.350.865.468.5\u2014TerminalBench (Best)66.5 (Claude Code)56.2(Claude Code)\u201457.0(Claude Code)46.4(Claude Code)\u2014\u2014\u201475.1(Codex)CyberGym68.748.3\u2014\u201417.341.366.6\u2014\u2014BrowseComp68.062.0\u2014\u201451.460.6\u2014\u2014\u2014BrowseComp (CM)79.375.9\u2014\u201467.674.984.085.982.7\u03c4\u00b3-Bench70.669.270.767.669.266.072.467.172.9MCP-Atlas71.869.274.148.862.263.873.869.267.2Tool-Decathlon40.738.039.846.335.227.847.248.854.6Vending Bench 2 ($)$5634$4432$5115\u2014$1034$1198$8018$911$6144\u041d\u043e \u043d\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0439\u0442\u0435 CUDA 13.2 \u0434\u043b\u044f GGUF, \u044d\u0442\u043e \u043f\u043e\u0440\u0442\u0438\u0442 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u043e\u0442\u0432\u0435\u0442\u043e\u0432.  \u0422\u0440\u0435\u0431\u043e\u0432\u0430\u043d\u0438\u044f \u043a \u043e\u0431\u043e\u0440\u0443\u0434\u043e\u0432\u0430\u043d\u0438\u044e\u0412\u0435\u0440\u0441\u0438\u044f UD-IQ2_M  \u043f\u043e\u0434\u0445\u043e\u0434\u0438\u0442 \u0434\u043b\u044f Mac \u0441 256 \u0413\u0411 \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u0438. \u0422\u0430\u043a\u0436\u0435 \u043e\u043d\u0430 \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043d\u0430 \u041f\u041a \u0441 \u043e\u0434\u043d\u043e\u0439 \u0432\u0438\u0434\u0435\u043e\u043a\u0430\u0440\u0442\u043e\u0439 \u043d\u0430 24 \u0413\u0411 \u0438 256 \u0413\u0411 \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u0438 \u0447\u0435\u0440\u0435\u0437 MoE offloading. \u0414\u043b\u044f 1-\u0431\u0438\u0442\u043d\u043e\u0439 \u0432\u0435\u0440\u0441\u0438\u0438 \u043d\u0443\u0436\u043d\u043e 220 \u0413\u0411 \u043f\u0430\u043c\u044f\u0442\u0438, \u0434\u043b\u044f 8-\u0431\u0438\u0442\u043d\u043e\u0439 805 \u0413\u0411.  \u041f\u0430\u0440\u0430\u043c\u0435\u0442\u0440Default Settings (Most Tasks)Terminal Benchtemperature1.00.7top_p0.951.0max new tokens13107216384\u0412 GLM-5.1 \u043f\u043e \u0443\u043c\u043e\u043b\u0447\u0430\u043d\u0438\u044e \u0432\u043a\u043b\u044e\u0447\u0435\u043d \u0440\u0435\u0436\u0438\u043c \u0440\u0430\u0441\u0441\u0443\u0436\u0434\u0435\u043d\u0438\u044f. \u0414\u043b\u044f \u0435\u0433\u043e \u043e\u0442\u043a\u043b\u044e\u0447\u0435\u043d\u0438\u044f \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0439\u0442\u0435—chat-template-kwargs ‘{«enable_thinking»:false}’\u0418\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 Unsloth Studio  \u0423\u0441\u0442\u0430\u043d\u043e\u0432\u043a\u0430 \u0434\u043b\u044f MacOS, Linux, WSLcurl -fsSL https:\/\/unsloth.ai\/install.sh | sh\u0423\u0441\u0442\u0430\u043d\u043e\u0432\u043a\u0430 \u0434\u043b\u044f Windows PowerShell irm https:\/\/unsloth.ai\/install.ps1 | iex\u0417\u0430\u043f\u0443\u0441\u043aunsloth studio -H 0.0.0.0 -p 8888\u041f\u043e\u0441\u043b\u0435 \u0437\u0430\u043f\u0443\u0441\u043a\u0430 \u043e\u0442\u043a\u0440\u043e\u0439\u0442\u0435 \u00a0http:\/\/localhost:8888   \u0412 \u043f\u043e\u0438\u0441\u043a\u0435 \u043d\u0430\u0439\u0434\u0438\u0442\u0435 GLM-5.1. \u0420\u0435\u043a\u043e\u043c\u0435\u043d\u0434\u0443\u0435\u0442\u0441\u044f \u0432\u044b\u0431\u0438\u0440\u0430\u0442\u044c \u0432\u0435\u0440\u0441\u0438\u044e UD-Q2_K_XL. \u0415\u0441\u043b\u0438 VRAM \u043d\u0435 \u0445\u0432\u0430\u0442\u0438\u0442, \u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u0437\u0430\u0434\u0435\u0439\u0441\u0442\u0432\u0443\u0435\u0442 \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u0443\u044e \u043f\u0430\u043c\u044f\u0442\u044c.\u0418\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 llama.cpp  \u0421\u043e\u0431\u0435\u0440\u0438\u0442\u0435 llama.cpp \u0438\u0437 \u0440\u0435\u043f\u043e\u0437\u0438\u0442\u043e\u0440\u0438\u044f GitHub. \u0414\u043b\u044f \u0441\u0438\u0441\u0442\u0435\u043c \u0431\u0435\u0437 GPU \u0438\u043b\u0438 \u043d\u0430 Mac \u0441 Metal \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0439\u0442\u0435\u00a0-DGGML_CUDA=OFF.    \u0414\u043b\u044f \u0441\u0431\u043e\u0440\u043a\u0438 apt-get updateapt-get install pciutils build-essential cmake curl libcurl4-openssl-dev -ygit clone https:\/\/github.com\/ggml-org\/llama.cppcmake llama.cpp -B llama.cpp\/build \\    -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ONcmake —build llama.cpp\/build —config Release -j —clean-first —target llama-cli llama-mtmd-cli llama-server llama-gguf-splitcp llama.cpp\/build\/bin\/llama-* llama.cpp  \u0417\u0430\u043f\u0443\u0441\u043a \u043e\u0431\u044b\u0447\u043d\u043e\u0439 \u0432\u0435\u0440\u0441\u0438\u0438 export LLAMA_CACHE=»unsloth\/GLM-5.1-GGUF».\/llama.cpp\/llama-cli \\    -hf unsloth\/GLM-5.1-GGUF:UD-IQ2_M \\    —ctx-size 16384 \\    —temp 0.7 \\    —top-p 1.0  \u0417\u0430\u043f\u0443\u0441\u043a \u0434\u043b\u044f \u0432\u044b\u0437\u043e\u0432\u0430 \u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442\u043e\u0432export LLAMA_CACHE=»unsloth\/GLM-5.1-GGUF».\/llama.cpp\/llama-cli \\    -hf unsloth\/GLM-5.1-GGUF:UD-IQ2_M \\    —ctx-size 16384 \\    —temp 1.0 \\    —top-p 0.95\u0420\u0430\u0431\u043e\u0442\u0430 \u0447\u0435\u0440\u0435\u0437 API  \u0414\u043b\u044f \u0437\u0430\u043f\u0443\u0441\u043a\u0430 \u0441\u0435\u0440\u0432\u0435\u0440\u0430 .\/llama.cpp\/llama-server \\    —model unsloth\/GLM-5.1-GGUF\/UD-IQ2_M\/GLM-5.1-UD-IQ2_M-00001-of-00006.gguf \\    —alias «unsloth\/GLM-5.1» \\    —prio 3 \\    —temp 1.0 \\    —top-p 0.95 \\    —ctx-size 16384 \\    —port 8001  \u041f\u0440\u0438\u043c\u0435\u0440 \u043a\u043e\u0434\u0430 \u043d\u0430 Python \u0434\u043b\u044f \u043e\u0431\u0440\u0430\u0449\u0435\u043d\u0438\u044f \u043a \u0441\u0435\u0440\u0432\u0435\u0440\u0443from openai import OpenAIopenai_client = OpenAI(    base_url = «http:\/\/127.0.0.1:8001\/v1»,    api_key = «sk-no-key-required»,)completion = openai_client.chat.completions.create(    model = «unsloth\/GLM-5.1»,    messages = [{«role»: «user», «content»: «Create a Snake game.»},],)print(completion.choices[0].message.content)\u0412\u044b\u0437\u043e\u0432 \u0444\u0443\u043d\u043a\u0446\u0438\u0439GLM-5.1 \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442 \u0432\u044b\u0437\u043e\u0432 \u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442\u043e\u0432. \u0414\u043b\u044f \u044d\u0442\u043e\u0433\u043e \u043d\u0443\u0436\u043d\u043e \u043e\u043f\u0440\u0435\u0434\u0435\u043b\u0438\u0442\u044c \u0444\u0443\u043d\u043a\u0446\u0438\u0438 \u0432 Python \u0438 \u043f\u0435\u0440\u0435\u0434\u0430\u0442\u044c \u0438\u0445 \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u0432 \u043c\u043e\u0434\u0435\u043b\u044c. def add_number(a: float | str, b: float | str) -> float:    return float(a) + float(b)def terminal(command: str) -> str:    if «rm» in command or «sudo» in command or «dd» in command or «chmod» in command:        return «Dangerous command»    return str(subprocess.run(command, capture_output = True, text = True, shell = True).stdout)\u041c\u043e\u0434\u0435\u043b\u044c \u043c\u043e\u0436\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u044d\u0442\u0438 \u0444\u0443\u043d\u043a\u0446\u0438\u0438 \u0434\u043b\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0440\u0430\u0441\u0447\u0435\u0442\u043e\u0432 \u0438\u043b\u0438 \u043a\u043e\u043c\u0430\u043d\u0434 \u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u0435. \u041f\u0440\u0438 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0438\u00a0llama-server\u00a0\u0430\u0432\u0442\u043e\u043c\u0430\u0442\u0438\u0447\u0435\u0441\u043a\u0430\u044f \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0430 \u0432\u044b\u0437\u043e\u0432\u043e\u0432 \u043d\u0430\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0435\u0442\u0441\u044f \u0447\u0435\u0440\u0435\u0437 \u0446\u0438\u043a\u043b, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u043f\u0440\u043e\u0432\u0435\u0440\u044f\u0435\u0442 \u043d\u0430\u043b\u0438\u0447\u0438\u0435\u00a0tool_calls\u00a0\u0432 \u043e\u0442\u0432\u0435\u0442\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 \u0438 \u0432\u043e\u0437\u0432\u0440\u0430\u0449\u0430\u0435\u0442 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0444\u0443\u043d\u043a\u0446\u0438\u0438.\u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438 https:\/\/habr.com\/ru\/articles\/1022242\/<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-475496","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/475496","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=475496"}],"version-history":[{"count":0,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/475496\/revisions"}],"wp:attachment":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=475496"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=475496"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=475496"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}