להורדה | SparkMoE גרסה V0.4.2 - הרצת מודלי שפה גדולים על חומרה חלשה! עדכון י"א תשרי תשפ"ז

[טכנופוב]

@א.מ.ד. נכון להיום על איזה מודלי MoE אתה ממליץ לחומרה 16GB RAM, למטרות שיחה בעברית ידע כללי/קידוד.

[א.מ.ד.]

@טכנופוב כתב:

@א.מ.ד. נכון להיום על איזה מודלי MoE אתה ממליץ לחומרה 16GB RAM, למטרות שיחה בעברית ידע כללי/קידוד.

Qwen 3.6 35B לקוד וסוכן, ו-Gemma 4 26B לשיחה כללית בעברית וידע כללי.

[טכנופוב]

@א.מ.ד. כתב:

Qwen 3.6 35B

זה שוקל 71.9 GB, אין משהוא יותר קל?
עריכה: מצאתי קבצים לא רישמיים שעברו קוונטיזציה, השאלה האם התוכנה תומכת במודלי IQ4_XS או רק בQ4_K_M

[א.מ.ד.]

@טכנופוב יש קבצי קוונטיזציה רשמיים של Qwen עצמם, אבל הכי מומלצים זה של Unsloth.
בכימותים של 4 ביט, המודל שוקל בסביבות ה-24 GB.
כמובן, כל כימות נתמך.
בפוסט שחרור הגרסה הראשונה שמתי קישורים מדוייקים למודלים המומלצים.

[טכנופוב]

@א.מ.ד. הורדתי על פי המלצתך https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/blob/main/Qwen3.6-35B-A3B-UD-IQ4_NL.gguf ולמרות שזה מודל MTP התוכנה לא מזהה אותו ככזה

[א.מ.ד.]

@טכנופוב כתב:

@א.מ.ד. הורדתי על פי המלצתך https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/blob/main/Qwen3.6-35B-A3B-UD-IQ4_NL.gguf ולמרות שזה מודל MTP התוכנה לא מזהה אותו ככזה

תתעלם מזה.

[טכנופוב]

@א.מ.ד. הומ..

בדקתי SHA-256 ויצא תקין

[א.מ.ד.]

@טכנופוב תעלה את הלוג

[עידו300]

??

Spoiler

[2026-11-05 13:32:21] SparkMoE 0.4.2-alpha session started
[2026-11-05 13:32:21] SparkMoE: Notification area icon ready
0.00.010.474 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)
0.00.015.207 I srv load_models: Loaded 0 cached model presets
0.00.017.942 I srv load_models: Loaded 2 custom model presets from C:\Users\User\AppData\Local\SparkMoE\model-profiles.ini
0.00.018.342 I srv operator (): Available models (2) (: custom preset)
0.00.018.350 I srv operator (): * lfm3-1.7b-q4-k-m (aliases: LFM3-1.7B-Q4_K_M)
0.00.018.353 I srv operator (): * qwen3.6-35b-a3b-ud-iq4-nl (aliases: Qwen3.6-35B-A3B-UD-IQ4_NL)
0.00.018.583 W srv llama_server: -----------------
0.00.018.585 W srv llama_server: CORS is set to allow all origins ('') and no API key is set
0.00.018.586 W srv llama_server: this can be a security risk (cross-origin attacks)
0.00.018.588 W srv llama_server: more info: https://github.com/ggml-org/llama.cpp/pull/25655
0.00.018.590 W srv llama_server: -----------------
0.00.018.617 W srv llama_server: -----------------
0.00.018.619 W srv llama_server: the following feature(s) are enabled:
0.00.018.621 W srv llama_server: router mode
0.00.018.623 W srv llama_server: do not expose the server to untrusted environments
0.00.018.625 W srv llama_server: -----------------
0.00.018.648 I srv llama_server: starting server in router mode. models will be automatically loaded on-demand
0.00.024.982 I srv llama_server: listening on http://127.0.0.1:8080
0.00.024.991 W srv llama_server: NOTICE: server default port will be changed to :9931 in a future release
0.00.024.991 W srv llama_server: ref: https://github.com/ggml-org/llama.cpp/pull/26508
[2026-11-05 13:32:22] SparkMoE: desktop: navigation ready
[2026-11-05 13:32:36] SparkMoE: desktop: navigation ready
0.22.848.693 I srv load_models: Loaded 0 cached model presets
0.22.855.660 I srv load_models: Loaded 1 custom model presets from C:\Users\User\AppData\Local\SparkMoE\model-profiles.ini
0.22.855.744 I srv load_models: (reload) removing model name=qwen3.6-35b-a3b-ud-iq4-nl (no longer in source)
0.22.855.878 I srv operator (): Available models (1) (*: custom preset)
0.22.855.884 I srv operator (): * lfm3-1.7b-q4-k-m (aliases: LFM3-1.7B-Q4_K_M)
0.22.859.514 I srv load: spawning server instance with name=lfm3-1.7b-q4-k-m on port 51264
0.22.859.629 I srv load: spawning server instance with args:
0.22.859.630 I srv load: C:\Users\User\Downloads\sparkmoe-0.4.2-alpha-windows-x64-avx2\llama-server.exe
0.22.859.631 I srv load: --context-shift
0.22.859.634 I srv load: --host
0.22.859.636 I srv load: 127.0.0.1
0.22.859.637 I srv load: --log-colors
0.22.859.638 I srv load: off
0.22.859.639 I srv load: --log-timestamps
0.22.859.640 I srv load: --metrics
0.22.859.641 I srv load: --moe-cache-mib
0.22.859.642 I srv load: 3885
0.22.859.643 I srv load: --moe-cache-policy
0.22.859.643 I srv load: lru
0.22.859.645 I srv load: --moe-dense-residency
0.22.859.645 I srv load: auto
0.22.859.646 I srv load: --moe-io-mode
0.22.859.647 I srv load: buffered
0.22.859.648 I srv load: --moe-io-threads
0.22.859.648 I srv load: 1
0.22.859.649 I srv load: --moe-overlap
0.22.859.650 I srv load: off
0.22.859.651 I srv load: --moe-paged-layers
0.22.859.651 I srv load: 0
0.22.859.653 I srv load: --moe-paging
0.22.859.653 I srv load: auto
0.22.859.654 I srv load: --moe-ple-cache-mib
0.22.859.655 I srv load: 64
0.22.859.655 I srv load: --moe-prefetch
0.22.859.666 I srv load: off
0.22.859.667 I srv load: --port
0.22.859.668 I srv load: 51264
0.22.859.669 I srv load: --spec-type
0.22.859.670 I srv load: none
0.22.859.670 I srv load: --alias
0.22.859.671 I srv load: lfm3-1.7b-q4-k-m
0.22.859.673 I srv load: --batch-size
0.22.859.674 I srv load: 1024
0.22.859.675 I srv load: --ctx-size
0.22.859.676 I srv load: 8192
0.22.859.677 I srv load: --cache-ram
0.22.859.677 I srv load: 0
0.22.859.678 I srv load: --cache-type-k
0.22.859.679 I srv load: f16
0.22.859.680 I srv load: --cache-type-v
0.22.859.681 I srv load: f16
0.22.859.682 I srv load: --swa-checkpoints
0.22.859.683 I srv load: 4
0.22.859.684 I srv load: --model
0.22.859.685 I srv load: C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf
0.22.859.685 I srv load: --n-gpu-layers
0.22.859.686 I srv load: 0
0.22.859.687 I srv load: --parallel
0.22.859.687 I srv load: 1
0.22.859.688 I srv load: --reasoning
0.22.859.689 I srv load: auto
0.22.859.690 I srv load: --ubatch-size
0.22.859.691 I srv load: 84
[51264] warning: no usable GPU found, --gpu-layers option will be ignored

[51264] warning: one possible reason is that llama.cpp was compiled without GPU support

[51264] warning: consult docs/build.md for compilation instructions

[51264] 0.00.010.942 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)

[51264] 0.00.011.462 W srv llama_server: -----------------

[51264] 0.00.011.467 W srv llama_server: CORS is set to allow all origins ('*') and no API key is set

[51264] 0.00.011.468 W srv llama_server: this can be a security risk (cross-origin attacks)

[51264] 0.00.011.469 W srv llama_server: more info: https://github.com/ggml-org/llama.cpp/pull/25655

[51264] 0.00.011.470 W srv llama_server: -----------------

[51264] 0.00.020.172 I srv load_model: loading model 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51264] 0.00.020.187 I MoE paging auto selected explicit mode because a cache limit was provided

[51264] 0.00.347.346 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51264] 0.00.347.354 E llama_model_load_from_file_impl: failed to load model

[51264] 0.00.347.414 E common_fit_params: encountered an error while trying to fit params to free device memory: failed to load model

[51264] 0.00.614.568 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51264] 0.00.614.579 E llama_model_load_from_file_impl: failed to load model

[51264] 0.00.614.582 W cmn common_init_: MoE paging auto could not use explicit expert paging; retrying with standard mmap loading

[51264] 0.00.905.861 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51264] 0.00.905.994 E llama_model_load_from_file_impl: failed to load model

[51264] 0.00.906.003 E cmn common_init_: failed to load model 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51264] 0.00.906.022 E srv load_model: failed to load model, 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51264] 0.00.906.031 I srv operator (): operator (): cleaning up before exit...

[51264] 0.00.906.681 E srv llama_server: exiting due to model loading error

0.23.835.161 I srv operator (): instance name=lfm3-1.7b-q4-k-m exited with status 1
0.27.648.072 I srv load_models: Loaded 0 cached model presets
0.27.661.480 I srv load_models: Loaded 1 custom model presets from C:\Users\User\AppData\Local\SparkMoE\model-profiles.ini
0.27.661.693 I srv operator (): Available models (1) (*: custom preset)
0.27.661.711 I srv operator (): * lfm3-1.7b-q4-k-m (aliases: LFM3-1.7B-Q4_K_M)
0.27.663.948 I srv load: spawning server instance with name=lfm3-1.7b-q4-k-m on port 51277
0.27.664.006 I srv load: spawning server instance with args:
0.27.664.011 I srv load: C:\Users\User\Downloads\sparkmoe-0.4.2-alpha-windows-x64-avx2\llama-server.exe
0.27.664.012 I srv load: --context-shift
0.27.664.013 I srv load: --host
0.27.664.014 I srv load: 127.0.0.1
0.27.664.015 I srv load: --log-colors
0.27.664.015 I srv load: off
0.27.664.019 I srv load: --log-timestamps
0.27.664.020 I srv load: --metrics
0.27.664.021 I srv load: --moe-cache-mib
0.27.664.022 I srv load: 3880
0.27.664.024 I srv load: --moe-cache-policy
0.27.664.024 I srv load: lru
0.27.664.025 I srv load: --moe-dense-residency
0.27.664.026 I srv load: auto
0.27.664.026 I srv load: --moe-io-mode
0.27.664.027 I srv load: buffered
0.27.664.028 I srv load: --moe-io-threads
0.27.664.028 I srv load: 1
0.27.664.029 I srv load: --moe-overlap
0.27.664.030 I srv load: off
0.27.664.030 I srv load: --moe-paged-layers
0.27.664.031 I srv load: 0
0.27.664.032 I srv load: --moe-paging
0.27.664.033 I srv load: auto
0.27.664.033 I srv load: --moe-ple-cache-mib
0.27.664.034 I srv load: 64
0.27.664.035 I srv load: --moe-prefetch
0.27.664.035 I srv load: off
0.27.664.036 I srv load: --port
0.27.664.037 I srv load: 51277
0.27.664.037 I srv load: --spec-type
0.27.664.038 I srv load: none
0.27.664.039 I srv load: --alias
0.27.664.039 I srv load: lfm3-1.7b-q4-k-m
0.27.664.040 I srv load: --batch-size
0.27.664.041 I srv load: 1024
0.27.664.041 I srv load: --ctx-size
0.27.664.045 I srv load: 8192
0.27.664.046 I srv load: --cache-ram
0.27.664.047 I srv load: 0
0.27.664.047 I srv load: --cache-type-k
0.27.664.048 I srv load: f16
0.27.664.049 I srv load: --cache-type-v
0.27.664.049 I srv load: f16
0.27.664.050 I srv load: --swa-checkpoints
0.27.664.050 I srv load: 4
0.27.664.051 I srv load: --model
0.27.664.052 I srv load: C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf
0.27.664.053 I srv load: --n-gpu-layers
0.27.664.053 I srv load: 0
0.27.664.054 I srv load: --parallel
0.27.664.055 I srv load: 1
0.27.664.055 I srv load: --reasoning
0.27.664.056 I srv load: auto
0.27.664.057 I srv load: --ubatch-size
0.27.664.057 I srv load: 84
[51277] warning: no usable GPU found, --gpu-layers option will be ignored

[51277] warning: one possible reason is that llama.cpp was compiled without GPU support

[51277] warning: consult docs/build.md for compilation instructions

[51277] 0.00.011.508 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)

[51277] 0.00.012.083 W srv llama_server: -----------------

[51277] 0.00.012.087 W srv llama_server: CORS is set to allow all origins ('*') and no API key is set

[51277] 0.00.012.088 W srv llama_server: this can be a security risk (cross-origin attacks)

[51277] 0.00.012.088 W srv llama_server: more info: https://github.com/ggml-org/llama.cpp/pull/25655

[51277] 0.00.012.088 W srv llama_server: -----------------

[51277] 0.00.020.710 I srv load_model: loading model 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51277] 0.00.020.728 I MoE paging auto selected explicit mode because a cache limit was provided

[51277] 0.00.314.188 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51277] 0.00.314.198 E llama_model_load_from_file_impl: failed to load model

[51277] 0.00.314.266 E common_fit_params: encountered an error while trying to fit params to free device memory: failed to load model

[51277] 0.00.612.703 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51277] 0.00.612.716 E llama_model_load_from_file_impl: failed to load model

[51277] 0.00.612.720 W cmn common_init_: MoE paging auto could not use explicit expert paging; retrying with standard mmap loading

[51277] 0.00.890.146 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51277] 0.00.890.155 E llama_model_load_from_file_impl: failed to load model

[51277] 0.00.890.159 E cmn common_init_: failed to load model 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51277] 0.00.890.172 E srv load_model: failed to load model, 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51277] 0.00.890.179 I srv operator (): operator (): cleaning up before exit...

[51277] 0.00.890.913 E srv llama_server: exiting due to model loading error

0.28.628.314 I srv operator (): instance name=lfm3-1.7b-q4-k-m exited with status 1
[2026-11-05 13:33:41] SparkMoE: desktop: navigation failed HRESULT=0x800704d4
1.28.773.310 I srv load_models: Loaded 0 cached model presets
1.28.788.371 I srv load_models: Loaded 1 custom model presets from C:\Users\User\AppData\Local\SparkMoE\model-profiles.ini
1.28.788.600 I srv operator (): Available models (1) (*: custom preset)
1.28.788.609 I srv operator (): * lfm3-1.7b-q4-k-m (aliases: LFM3-1.7B-Q4_K_M)
1.28.792.378 I srv load: spawning server instance with name=lfm3-1.7b-q4-k-m on port 51334
1.28.792.459 I srv load: spawning server instance with args:
1.28.792.460 I srv load: C:\Users\User\Downloads\sparkmoe-0.4.2-alpha-windows-x64-avx2\llama-server.exe
1.28.792.461 I srv load: --context-shift
1.28.792.462 I srv load: --host
1.28.792.462 I srv load: 127.0.0.1
1.28.792.463 I srv load: --log-colors
1.28.792.464 I srv load: off
1.28.792.465 I srv load: --log-timestamps
1.28.792.467 I srv load: --metrics
1.28.792.469 I srv load: --moe-cache-mib
1.28.792.470 I srv load: 4150
1.28.792.472 I srv load: --moe-cache-policy
1.28.792.474 I srv load: lru
1.28.792.476 I srv load: --moe-dense-residency
1.28.792.478 I srv load: auto
1.28.792.479 I srv load: --moe-io-mode
1.28.792.481 I srv load: buffered
1.28.792.483 I srv load: --moe-io-threads
1.28.792.485 I srv load: 1
1.28.792.487 I srv load: --moe-overlap
1.28.792.488 I srv load: off
1.28.792.490 I srv load: --moe-paged-layers
1.28.792.492 I srv load: 0
1.28.792.494 I srv load: --moe-paging
1.28.792.496 I srv load: auto
1.28.792.498 I srv load: --moe-ple-cache-mib
1.28.792.500 I srv load: 64
1.28.792.502 I srv load: --moe-prefetch
1.28.792.503 I srv load: off
1.28.792.506 I srv load: --port
1.28.792.507 I srv load: 51334
1.28.792.509 I srv load: --spec-type
1.28.792.511 I srv load: none
1.28.792.512 I srv load: --alias
1.28.792.513 I srv load: lfm3-1.7b-q4-k-m
1.28.792.514 I srv load: --batch-size
1.28.792.515 I srv load: 1024
1.28.792.516 I srv load: --ctx-size
1.28.792.517 I srv load: 8192
1.28.792.518 I srv load: --cache-ram
1.28.792.519 I srv load: 0
1.28.792.520 I srv load: --cache-type-k
1.28.792.520 I srv load: f16
1.28.792.521 I srv load: --cache-type-v
1.28.792.522 I srv load: f16
1.28.792.523 I srv load: --swa-checkpoints
1.28.792.523 I srv load: 4
1.28.792.524 I srv load: --model
1.28.792.525 I srv load: C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf
1.28.792.526 I srv load: --n-gpu-layers
1.28.792.526 I srv load: 0
1.28.792.527 I srv load: --parallel
1.28.792.528 I srv load: 1
1.28.792.529 I srv load: --reasoning
1.28.792.529 I srv load: auto
1.28.792.530 I srv load: --ubatch-size
1.28.792.531 I srv load: 88
1.28.809.798 I srv ensure_model: waiting until model name=lfm3-1.7b-q4-k-m is fully loaded...
[51334] warning: no usable GPU found, --gpu-layers option will be ignored

[51334] warning: one possible reason is that llama.cpp was compiled without GPU support

[51334] warning: consult docs/build.md for compilation instructions

[51334] 0.00.012.767 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)

[51334] 0.00.013.309 W srv llama_server: -----------------

[51334] 0.00.013.314 W srv llama_server: CORS is set to allow all origins ('*') and no API key is set

[51334] 0.00.013.314 W srv llama_server: this can be a security risk (cross-origin attacks)

[51334] 0.00.013.315 W srv llama_server: more info: https://github.com/ggml-org/llama.cpp/pull/25655

[51334] 0.00.013.315 W srv llama_server: -----------------

[51334] 0.00.019.311 I srv load_model: loading model 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51334] 0.00.019.330 I MoE paging auto selected explicit mode because a cache limit was provided

[51334] 0.00.361.572 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51334] 0.00.361.593 E llama_model_load_from_file_impl: failed to load model

[51334] 0.00.361.662 E common_fit_params: encountered an error while trying to fit params to free device memory: failed to load model

[51334] 0.00.695.686 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51334] 0.00.695.767 E llama_model_load_from_file_impl: failed to load model

[51334] 0.00.695.809 W cmn common_init_: MoE paging auto could not use explicit expert paging; retrying with standard mmap loading

[51334] 0.01.027.943 E llama_model_load: error loading model: unknown model architecture: 'lfm3'

[51334] 0.01.027.957 E llama_model_load_from_file_impl: failed to load model

[51334] 0.01.027.961 E cmn common_init_: failed to load model 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51334] 0.01.027.978 E srv load_model: failed to load model, 'C:\Users\User\AppData\Local\SparkMoE\models\download-1\LFM3-1.7B-Q4_K_M.gguf'

[51334] 0.01.027.987 I srv operator (): operator (): cleaning up before exit...

[51334] 0.01.028.651 E srv llama_server: exiting due to model loading error

1.29.891.377 I srv operator (): instance name=lfm3-1.7b-q4-k-m exited with status 1
1.29.891.661 W srv operator (): got exception: {"error":{"code":500,"message":"model name=lfm3-1.7b-q4-k-m failed to load","type":"server_error"}}

[א.מ.ד.]

@עידו300 זה בגלל שהמודל הזה לא נתמך ב-SparkMoE בגרסה הנוכחית.
בגרסה הבאה של SparkMoE בעז"ה יהיה עדכון אוטומטי של מנוע llama.cpp.