Advisor / Hardware / AMD Ryzen AI Max+ 395 (Strix Halo) 128GB

Best local LLMs for AMD Ryzen AI Max+ 395 (Strix Halo) 128GB

AMD Ryzen AI Max+ 395 (Strix Halo) 128GB. Memory-fit math with KV cache, speeds labelled by how they were obtained, and commands you can paste.

108.3 GBusable for model + KV cache
256 GB/smemory bandwidth
26measured data points on file
Sponsor slot. No sponsor yet and no link.

Best for Coding · 16k context

ModelQuantFitMemorySpeed (generation)
gpt-oss 120BQ8_0fits fully64.0 GB~44 tok/s scaled from measured src
Qwen3-Coder 30B-A3BQ8_0fits fully34.1 GB~55 tok/s scaled from measured src
Qwen3.6 35B-A3BQ8_0fits fully37.3 GB~53 tok/s measured src
Gemma 4 26B-A4BQ8_0fits fully27.9 GB~35 tok/s scaled from measured src
gpt-oss 20BQ8_0fits fully12.5 GB~93 tok/s scaled from measured src
llama.cpp · top pick
llama-server -hf unsloth/gpt-oss-120b-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
ollama · top pick
OLLAMA_CONTEXT_LENGTH=16384 ollama run hf.co/unsloth/gpt-oss-120b-GGUF:Q8_0

Best for Chat / general · 16k context

ModelQuantFitMemorySpeed (generation)
gpt-oss 120BQ8_0fits fully64.0 GB~44 tok/s scaled from measured src
Qwen3.6 35B-A3BQ8_0fits fully37.3 GB~53 tok/s measured src
Qwen3-Coder 30B-A3BQ8_0fits fully34.1 GB~55 tok/s scaled from measured src
Gemma 4 26B-A4BQ8_0fits fully27.9 GB~35 tok/s scaled from measured src
gpt-oss 20BQ8_0fits fully12.5 GB~93 tok/s scaled from measured src
llama.cpp · top pick
llama-server -hf unsloth/gpt-oss-120b-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
ollama · top pick
OLLAMA_CONTEXT_LENGTH=16384 ollama run hf.co/unsloth/gpt-oss-120b-GGUF:Q8_0

Best for Vision · 16k context

ModelQuantFitMemorySpeed (generation)
Qwen3.6 35B-A3BQ8_0fits fully37.3 GB~53 tok/s measured src
Gemma 4 26B-A4BQ8_0fits fully27.9 GB~35 tok/s scaled from measured src
Qwen3.5 4BQ8_0fits fully5.1 GB~23 tok/s scaled from measured src
Gemma 4 31BQ4_K_Mfits fully20.5 GB~10 tok/s measured src
Qwen3.6 27BQ4_K_Mfits fully18.1 GB~13 tok/s measured src
llama.cpp · top pick
llama-server -hf unsloth/Qwen3.6-35B-A3B-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
ollama · top pick
OLLAMA_CONTEXT_LENGTH=16384 ollama run hf.co/unsloth/Qwen3.6-35B-A3B-GGUF:Q8_0

All 13 models that fit, fastest first · 16k context

Click a column header to sort. 13 of 13 speeds here are measured or scaled from measured; the rest are bandwidth estimates. Filter by use case, speed source, fit, license and more in the interactive advisor →

ModelParamsQuantFitMemoryTok/sSource
gpt-oss 20B20.91B MoEQ8_0fits fully12.5 GB~93scaled from measured
Qwen3-Coder 30B-A3B30.53B MoEQ8_0fits fully34.1 GB~55scaled from measured
Qwen3.6 35B-A3B35.95B MoEQ8_0fits fully37.3 GB~53measured
gpt-oss 120B116.83B MoEQ8_0fits fully64.0 GB~44scaled from measured
Gemma 4 26B-A4B25.81B MoEQ8_0fits fully27.9 GB~35scaled from measured
Qwen3.5 4B4.66BQ8_0fits fully5.1 GB~23scaled from measured
GLM-4.5-Air110.47B MoEQ6_Kfits fully102.1 GB~16scaled from measured
Qwen3.5 9B9.65BQ5_K_Mfits fully7.2 GB~16scaled from measured
Qwen3-VL 8B8.77BQ6_Kfits fully9.1 GB~16scaled from measured
Gemma 4 12B11.96BQ4_K_Mfits fully7.7 GB~15scaled from measured
Mistral Small 3.2 24B24.01BQ4_K_Mfits fully17.0 GB~14scaled from measured
Qwen3.6 27B27.78BQ4_K_Mfits fully18.1 GB~13measured
Gemma 4 31B31.27BQ4_K_Mfits fully20.5 GB~10measured

Long context on this hardware · up to 1M tokens

Largest context whose KV cache still fits next to the weights (same placement as the 16k result, KV types f16 / q8_0 / q4_0). Models are capped at their verified maximum; extended means beyond the native maximum using YaRN, only listed where the model card documents it. Long context needs flash attention (-fa on, already in the commands) and a lot of memory; a quantized KV cache saves memory but can cost accuracy. Speed at depth is an estimate (penalty fitted to measured depth runs up to 128k).

Try it in the advisor: 128k · 256k · 512k · 1M

ModelNative maxMax ctx f16Max ctx q8_0Max ctx q4_0Est. tok/s at max f16 ctx
Qwen3.6 35B-A3B Q8_0256k1M extended1M extended1M extended~9 estimated
Qwen3.5 4B Q8_0256k1M extended1M extended1M extended~4 estimated
Qwen3.5 9B Q5_K_M256k1M extended1M extended1M extended~3 estimated
Qwen3.6 27B Q4_K_M256k1M extended1M extended1M extended~4 estimated
Qwen3-Coder 30B-A3B Q8_0256k512k extended1M extended1M extended~5 estimated
Qwen3-VL 8B Q6_K256k512k extended1M extended1M extended~2 estimated
Gemma 4 26B-A4B Q8_0256k256k256k256k~18 estimated
Gemma 4 12B Q4_K_M256k256k256k256k~10 estimated
Gemma 4 31B Q4_K_M256k256k256k256k~6 estimated
gpt-oss 20B Q8_0128k128k128k128k~44 estimated
gpt-oss 120B Q8_0128k128k128k128k~20 estimated
Mistral Small 3.2 24B Q4_K_M128k128k128k128k~7 estimated
GLM-4.5-Air Q6_K128k32k64k128k~11 estimated
Affiliate box: “cheapest upgrade” for AMD Ryzen AI Max+ 395 (Strix Halo) 128GB. No affiliate links yet.

Open the interactive advisor with this hardware →

Speeds marked measured are medians of real runs (short-context generation) from these sources; scaled from measured means a measurement of a same-size or same-model-different-quant run, scaled by bytes read; estimated comes from memory bandwidth (spec source: https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html) × an efficiency factor of 0.80 calibrated to measured runs on this hardware. 26 measured data points on file for this hardware · submit your benchmark. Memory size on this page is an example configuration; check what your machine has. Model data last verified 2026-10-14. Ranking = fits and fast enough first, then larger; not a quality benchmark.

Page generated 2026-10-16 by tools/generate_pages.js