Setups with AMD Strix Halo (Ryzen AI Max+ 395): local inference speed (tok/s), sourced
As of 2026-11-14, this dataset has 26 recipes for AMD Strix Halo (Ryzen AI Max+ 395). 10 were measured by the poster. Fastest measured single-device, single-stream text result: 78 tok/s with LFM3-1.7B on llama.cpp (AMD Strix Halo iGPU; Measured, 2026-11-01; source). 20 entries have a date, from 2026-08-26 to 2026-11-09.
Filter these in the explorer →
| Model | Engine | Hardware | tok/s | Metric | Source level | Date | Source |
|---|---|---|---|---|---|---|---|
| LFM3-1.7B | llama.cpp | AMD Strix Halo iGPU | 78 | ROCm: decode about 78, prefill 4.1K; 7900 XTX + 7900 XT tensor split about 151 / about 7.1K | Measured | 2026-11-01 | link |
| Qwen3.8-Flash-Next | Halogen | AMD Radeon 8060S iGPU (Strix Halo, 128GB unified memory) | 66 | Halogen 0.18.22, 44 agent turns: decode mean 66, peak 89; cold prefill 1,135; prompt cache hits 99.4%; average context about 91K; MTP 3.5 tokens per round | Measured | 2026-11-03 | link |
| Qwen3.8-Flash-Next | llama.cpp (vahpetr fork) | AMD Strix Halo 128GB (Ryzen AI Max+ 395, Vulkan) | 61 | Empty context 61 t/s, 524K 45, 1M 35.5; prefill 8K 550 / 213K ~420; 1M context uses 102.7 GiB (stock llama.cpp on the same machine ~24–27) | Measured | 2026-10-10 | link |
| Qwen3.8-Flash-Next | Strata | AMD Strix Halo 395 | 47.7 | 8K prompt/128 output: 47.7 tok/s decode, 699 tok/s prefill, TTFT 11.8 s (halogen 0.17.2: 34.3 / 1,573; llama.cpp HIP: 23.2 / 466) | Measured | 2026-10-08 | link |
| Qwen3.8-Flash-Next | Halogen | Strix Halo 128GB | 38 | 38.0 tok/s average (old llama.cpp 13.9) | Measured | link | |
| Qwen3.8-Flash-Next | Halogen | AMD Strix Halo 395 | 34.3 | 34.3 tok/s decode, 1,573 tok/s prefill, TTFT 5.2 s (8K prompt) | Measured | 2026-10-08 | link |
| DeepSeek-V4.1-Flash-1017 | gufo | 1× Bosgame M5 (Strix Halo, 128GB) | 23.3 | Greedy, measured over HTTP: depth 0 / 32K / 128K: prefill pp2048 1,077 / 993 / 700; decode 23.3 / 22.9 / 21.7; DSpark on mixed text 28.0 / 30.1 / 27.2; DSpark on repetitive text (separate, reads high) 65.3 / 59.6 / 55.2; aggregate (separate) 8 users 78.5; editing a source file 67.6. tok_s is decode at depth 0 | Measured | 2026-10-27 | link |
| GLM-5.3-Flash | Unspecified | Strix Halo (Framework Desktop 395+) | 17 | 17–18 tok/s | Measured | link | |
| DeepSeek-V4.1-Flash | llama.cpp | ACEMAGIC Strix Halo 128GB (€2,300) + fast SSD | 15 | >15 tok/s, peak 31.7 | Measured | link | |
| DeepSeek-V4.1-Flash | Strata | AMD Ryzen AI Max+ 395 128GB (Strix Halo, Windows) | 15 | Code decode 15–21; prefill 214–233 at 64K context | Measured | 2026-10-29 | link |
| Qwen3.8-Flash-Next | Strata | Strix Halo + Radeon AI PRO R9700 | 96 | Strata 0.2.11 + Unsloth Dynamic 3.1: single-stream 96 at 4×256k context. Author reports a clear gain vs the prior quant. Note: Dynamic file revisions are not cross-backend comparable. tok_s uses 96 | Claimed | 2026-11-06 | link |
| DeepSeek-V4.1-Flash-1017 | Project Maya | AMD Strix Halo (Ryzen AI Max+ 395, 128GB) + 2× RX 7900 XTX (OCuLink) | 82.5 | Maya-XS v2: decode 82.5; prefill for a 491K prompt 1,195.5 → 1,448.2 (the sparse-attention indexer now uses RDNA 3 WMMA). Maya-S: decode 41.7, prefill at 130K 1,378 | Claimed | 2026-11-03 | link |
| Qwen3.8-Flash-Next | gufo | Strix Halo (model and memory not stated) | 73.17 | single user 73.17, pp 1,801; aggregate (separate) 8 users 194.37; Qwen3.8-27B Q4_K_XL + DFlash2 on the same engine 70.56 | Claimed | 2026-10-26 | link |
| Qwen3.8-27B | llama.cpp | AMD Strix Halo | 65.6 | Up to 65.6 tok/s decode | Relayed | 2026-08-26 | link |
| Qwen3.8-Flash-Next | Halowin Bubo | AMD Strix Halo laptop (70W) | 57.1 | Bubo 0.1.2: prose 57.1, agent 66.2, code edits 109; 32K prefill 1,639 (fast mode). Same file compared with gufo. tok_s uses prose 57.1 | Claimed | 2026-11-05 | link |
| DeepSeek-V4-Flash (V4, not 4.1) | Unspecified | Strix Halo + R9700 | 51.1 | vendor homepage 86 tok/s; independent retest 51–57 (median 51.1) | Claimed | link | |
| Qwen3.8-Flash-Next | Unspecified | AMD Strix Halo 128GB mini PC | 44 | 44–59 tok/s; ~1,400 PP | Relayed | link | |
| Qwen3.8-Flash-Next | vLLM | AMD Strix Halo 128GB | 43 | Single stream on real text: 5.6K prompt prefill 1,539 / decode 43; 26.9K 1,617 / 43; 100.5K 1,572 / 40. Before this work, prefill 684–806 and decode about 18. KLD vs stock 0.008–0.019, GSM8K 93 vs 92 | Claimed | 2026-10-30 | link |
| Qwen3.8-Flash-Next | gufo | Strix Halo laptop (gfx1151, 128GB; 96GB to GPU) + Windows 11 | 41 | gufo v0.31.0 plus five local patches (gist). Two sessions × 262144, q8_0 KV. Long-reply decode about 41 at 55–67K, 36–48 at 230K; cold prefill about 1240–1335; MTP acceptance 70–91%. No raw results; approximate; claim. Launch .bat in the post. | Claimed | 2026-11-09 | link |
| DeepSeek-V4.1-Flash-1017 | libosmos | AMD Strix Halo (Ryzen AI Max+ 395, 128GB) | 38 | Engine author: up to 38 with drafting, prefill >450 (Reddit post says >480). tok_s uses 38. 2026-11-09 libosmos v0.5.9 on the same Strix Halo 128GB: ud-iq1_s plain/drafted 19.5/42.5; ud-iq1_m 17.5/35.4; ud-iq2_xxs 14.4/26.5. Context omitted. Drafted 42.5 is above this row's peak 38, but without context the headline stays 38. | Claimed | 2026-11-05 | link |
| DeepSeek-V4.1-Flash-1017 | Unspecified | 2× AMD Strix Halo (USB4 link) | 35.3 | Decode 35.3; 41.9–70.6 with their own speculative drafter; prefill about 1,100 (8K–45K); aggregate (separate) 8 requests about 114 | Claimed | 2026-11-02 | link |
| Qwen3.6-35B-A3B | vLLM | AMD Strix Halo (Radeon 8060S, 128GB unified memory) | 32.6 | Stock vLLM → with libstrix: decode at 2K 17.8 → 32.6, at 64K 12.8 → 26.9; prefill at 2K 1,583 → 3,581, at 64K 260 → 2,617; a 256K prompt prefill 7,568 s → 241 s. 27B bf16 decode 4.3 → 4.4 (no gain) | Claimed | 2026-10-29 | link |
| Qwen3.8-Flash-Next-Gyro | llama.cpp (agentionai fork) | AMD Strix Halo | 31.7 | 31.7 tok/s decode | Claimed | 2026-09-30 | link |
| Laguna S 2.1 | gufo (mikealanni fork) | Ryzen AI MAX+ 395 (Strix Halo) | 25.9 | Decode 25.9; DFlash code 44–45, prose about 28; prefill 1,145 (GPU), 1,277 (GPU + NPU) | Claimed | 2026-11-01 | link |
| Comet-60B-A12B | llama.cpp | AMD Ryzen AI Max+ 395 / Radeon 8060S (Strix Halo) | 24.59 | Quant-author model card, llama.cpp b12223 Vulkan. Headline is llama-bench default parameters (-fa 1, no draft) tg128: 24.59; pp512 is 466.70. llama-server --spec-type draft-mtp, greedy, short coding request (417 generated tokens) twice 36.5 / 35.9 (87.5% draft accept) is a code figure and is not the headline. The server command does not set -c; context is left null (defaults). No raw log; claim. | Claimed | 2026-11-09 | link |
| Qwen3.8-27B | Unspecified | AMD R9700 32GB / Strix Halo | 159 tok/s / 64 tok/s | Relayed | link |
Numbers come from each source. This project did not measure them again. Models and settings differ between rows. Source levels: Official = project or vendor; Measured = the poster ran it (an engine or recipe author must also publish raw data and reproduction steps); Relayed = second-hand; Claimed = prediction or no reproduction details.