Server 120 — модели и фактические тесты

Ryzen AI Max+ 395 · Radeon 8060S / Strix Halo · gfx1151 · 128 GB UMA, рабочее окно около 96 GiB · Ubuntu 24.04 · обновлено 05.09.2026

Последний production baseline: Qwen3.6-35B-A3B Q4_K_M; сервис остановлен на время acceptance-серии. Все остальные модели тестировались изолированно. Этот каталог показывает не рекламные цифры, а локальные измерения и ограничения.

45–60 tok/sQwen3.6 short decode baseline
50/50AgentWorld sandbox loop
60.18 tok/sNemotron Omni vision decode
0/3Nemotron early-context retention
10.29 tok/sFlash-Next на 245K context
0Production side-effects in isolated tests

Быстрый выбор

ЗадачаМодельПочемуОграничение
Общий production text/agentQwen3.6-35B-A3B Q4_K_MЛучший подтверждённый баланс RU/EN, tools и скоростиСледовать production profile; не менять sampler без A/B
Sandbox tool/repoQwen-AgentWorld Q4_K_M50/50 tool loop, recovery, repo-edit, security PASSТолько безопасный sandbox
Fast helper/routerGranite 4.2-8B / LFM2.538–39 tok/s / 211–221 JSON tok/sVerifier обязателен; не primary model
Vision stagingNemotron Omni Q410/10 image transport, 60.18 decode tok/sGrounding 1/4, video abort
Research long contextFlash-Next MTPРабота до 245K, MTP +18–20% на коротком A/BCoding/strict output fail; production reject

Проверенные конфигурации

Общие правила запуска
HSA_USE_SVM=0
HSA_ENABLE_SDMA=0
llama-server -m MODEL.gguf --host 127.0.0.1 --port PORT \
  --ctx-size CONTEXT --parallel 1 --gpu-layers 999 \
  --jinja --flash-attn auto --seed 42
  • Тяжёлые модели запускать последовательно, не одновременно на UMA.
  • Для production Qwen3.6 закреплены --reasoning off --flash-attn off --no-mmap.
  • Для tools: строковый tool_choice:"required", enable_thinking:false, достаточный max_tokens.
  • MTP с parallel>1 считать недопустимым до отдельного slot-isolation gate.
  • Фактический context считать по /tokenize и usage.prompt_tokens.

Основные модели

Qwen3.6-35B-A3B Q4_K_M — 20.42 GB — PRODUCTION BASELINE

Runtime: b10276 ROCm/HIP gfx1151, -ngl 999, parallel=1; production: reasoning off, flash-attn off, no-mmap.

Speed: 45–60 tok/s short decode; control 58.83–59.54. Tools: corrected 15/15 short/8K/30K. Business: selection 10/10, exact args 9/10; date retry 5/5.

Recommendation: retain as general production model.

Qwen3.8-Flash-Next UD-IQ4_XS — ~87.2 GB + MTP 4.14 GB — RESEARCH / REJECT PRODUCTION

Runtime: EngramHalo ROCm, q8 KV, parallel=1, MTP draft Q8_0, n-max=4.

Speed: no MTP 19.49/17.83 tok/s at 8K/24K; MTP 23.06/21.39 (+18–20%). Long run ~10.29 tok/s at 245K, VRAM ~84 GB, soak 99/100.

Quality: 24K marker fails; coding usable content 0/12 or 1/12 depending budget, marker 3/45 maximum. Patched TOP_K/rocPRIM artifact absent. Existing MTP slot test 16/16 at parallel 1 and 16/16 at parallel 4 is build-specific.

IFM K2-Horizon-MoVA-36B-A4B Q4 — ~22.4 GB — REJECT GENERAL PRODUCTION

Separate MBZUAI-IFM fork; quality 6/10, tool/agent soak and serial stability PASS, but Russian/business quality insufficient. Q5/Q8 not tested after Q4 gate.

Nemotron-3-Nano-Omni-30B-A3B Q4 — 22.42 GB + mmproj 1.59 GB — MULTIMODAL EXPERIMENT

Runtime: b10276 ROCm, Reasoning-mmproj-F16, parallel=1. Text/JSON 220/220, stress 1000/1000, tools 15/15, recovery 5/5.

Text: context matrix 7,969/31,950/63,850 — exact 8K FAIL, 32K FAIL, 64K PASS; decode 63.3/59.5/54.3 tok/s.

Early context: token-counted 7,999/31,997/63,993 — 0/3, all returned 0123456789.

Vision: 10/10, wall 1.827 s, prefill 436.5, decode 60.18 tok/s, RAM 5.93 GiB, RSS 2.66 GiB, VRAM 23.6 GiB, GPU 97%. Grounding 1/4. Audio CLI PASS; video abort on patch_embeddings_1.

Qwen3.8-27B UD-Q5_K_XL — 20.88 GB — EXPERIMENTAL CODING SPECIALIST

b10276 ROCm, q8 KV, parallel=1. Coding 12/50; strict JSON 20/20; decode ~10.4 tok/s; soak 42/100; RU/business 36/45; tools 15/15.

32K recall 0/3. Token-counted 64K: 63,993 tokens, 303.9 s, prefill 211.5, decode 9.12 tok/s, strict recall FAIL. Full 50 coding/20 repo gate not complete.

Qwen-AgentWorld-35B-A3B Q4 — 22.13 GB — AGENT SPECIALIST

Vulkan b19cbe9, ctx 65K, q8 KV, parallel=1. Protocol, sandbox, recovery, MCP/web, repo-edit, security/rollback PASS. 50-turn: 50/50 за 93.177 s; security 7/7.

Использовать только в изолированном безопасном sandbox.

Ember + DeepSeek V4 Flash Vision Exp Q2 — 86.72 GB + encoder 0.93 GB — VISION RESEARCH

HSA_USE_SVM=0, HSA_ENABLE_SDMA=0; text 22.43 tok/s. Vision v2 8/10, markers 100%, JSON 80%; coding 1/10, agent 8/10, GUI 0/10, soak 22/100. OCR 4/4 markers and 3/4 JSON; format retry 8/10. Deferred, not production.

GLM-4.7-Flash Q4 — 18.13 GB — EXPERIMENTAL CANDIDATE

b10276 ROCm, ctx 16K, decode ~56.8 tok/s. RU smoke, arithmetic 5/5, tool 1/1, soak 20/20. JSON/code fenced in Markdown; strict output gate remains open. Q6/IQ4 file absent.

Granite 4.2-8B 5.35 GB / LFM2.5-1.2B 0.731 GB — FAST HELPERS

Granite: 38.2–38.9 tok/s, cold full response 4.886 s. LFM: ready 2.216 s, first response 2.407 s, 200 JSON about 211–221 tok/s. Both need verifier and are not production replacements.

Muse Glimmer 30B 16.76 GB + mmproj 1.40 GB — VISION STAGING

b10658 ROCm, text 12.6–12.9 tok/s; vision 5/5, 36.2–118.9 s/image; 1C smoke 194.4 s. Too slow for primary production.

Исторический и неполный inventory

Ранее тестированные модели
МодельРезультатСтатус
Qwen3.6 Q2 MTP39/42 quality, 100 sequentialExperimental
Qwen3-Coder-Next Q4~46.96 tok/sCoding candidate
Gemma 4 12B + mmproj25.5–28.18 tok/s, vision 3/3Vision specialist
Gemma 4 31B + mmproj10.6–11.51 tok/s, image PASSSlow vision
MiniCPM5-1B~216.7 tok/s targeted smoke; earlier 0/200Smoke-only
DeepSeek-R1-Distill-32B9.1–10.1 short; ~2.5 longReject interactive
GPT-OSS-120B MXFP4~50–51 tok/s compatible benchmark; pinned runtime failedRuntime-dependent
MiniMax-H3608×352 video 4/4; larger blockedVideo specialist
Установлены/упоминались, но acceptance нет

Qwen3-8B/14B/32B, Qwen2.5-Coder-32B, Qwen2.5-72B, Llama 3.3 70B, DeepSeek-R1-Distill-Llama-70B, LFM2.5-8B-A1B и Qwen3.8-2.4T-A95B. Metadata или наличие GGUF не означает локальный тест.

Полный дисковый snapshot моделей и проекторов

Snapshot каталога Server 120 от 03.09.2026; наличие файла не означает загрузку в VRAM или acceptance.

ФайлРазмерНазначение / статус
qwen2.5-0.5b-instruct-q4_k_m.gguf491 MBEmergency smoke/fallback
DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf1.12 GBSmall experimental
MiniCPM5-1B-Q4_K_M.gguf688 MBSmoke-plus, acceptance incomplete
LFM2.5-1.2B-Instruct-Q4_K_M.gguf731 MBFast preprocessor
mistral-7b-instruct-v0.2.Q4_K_M.gguf4.37 GBSaved candidate
Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf4.68 GBCoding candidate
gemma-4-12b-it-qat-q4_0.gguf + mmproj6.98 + 0.175 GBLight vision profile
T-lite-it-2.1-Q6_K.gguf6.72 GBCandidate
Qwen3-30B-A3B-Instruct-2507-Q4_K_M.gguf18.56 GBQwen candidate
Qwen3.6-27B-Q4_K_M.gguf19.10 GBDense balanced candidate
gemma-4-31B_q4_0-it.gguf + mmproj17.65 + 1.20 GBHigh-quality vision/manual
Qwen3.5-35B-A3B-Revised-q4_k_m.gguf21.17 GBExperimental Qwen
DeepSeek-R1-Distill-Qwen-32B-Q4_K_M.gguf19.85 GBSaved, service off
T-Search-Q5_K_M.gguf25.35 GBSearch/RAG candidate
T-pro-it-2.1-Q6_K.gguf26.88 GBCandidate
Nemotron Omni MXFP4 + mmproj21.73 + 1.59 GBHistorical multimodal experiment
Qwen3-Coder-Next-Q4_K_M.gguf48.53 GBSeparate coding runtime
gpt-oss-120b-MXFP4.gguf63.39 GBExperimental reasoning
Whisper large-v3 turbo/q5/q40.574–1.08 GBSpeech models; runtime acceptance blocked
Qwen3.6-35B-A3B-UD-Q2_K_XL.gguf12.57 GBLow-bit/MTP experimental
Qwen3.6-35B-A3B-Q4_K_M.gguf20.42 GBLast production baseline; service stopped for acceptance

Полный reconciliation: SERVER120_MODEL_CATALOG.md.

Открытые пункты

\n