{
 "generated": "2026-11-21",
 "method": "Measured results from published sources, not our own benchmarks. decode_tok_s is token generation speed; prefill_tok_s is prompt processing; ttft_s is time to first token. Context matters: a result at 4K is not a result at 128K. A missing prefill cell means the source did not report it, not that it is zero.",
 "rows": [
  {
   "id": "strata-win-iq3s",
   "model_id": "qwen3-8-flash-next-abliterated-gsq-rco-strata-gguf-gsq-rco",
   "model": "Qwen3.8 Flash-Next IQ3_S (Strata)",
   "machine": "GMKtec NucBox EVO-X2",
   "hardware": "Ryzen AI Max+ 395, 128 GB, Radeon 8060S",
   "os": "Windows 11 Pro",
   "engine": "Strata 0.2.7 / ROCm 10.2",
   "context": "short chat",
   "ttft_s": 0.6,
   "kind": "measured",
   "source": "https://github.com/Niko1221/Strata/blob/HEAD/docs/STRIX_HALO_WINDOWS.md",
   "note": "Third-party measurement on the closest M5-class box, not a Bosgame. 4K/16K/32K prompt reads were 755/821/792 tok/s; 4K TTFT 0.60 s.",
   "decode_tok_s": 44.0,
   "prefill_tok_s": 755
  },
  {
   "id": "strata-win-code",
   "model_id": "qwen3-8-flash-next-abliterated-gsq-rco-strata-gguf-gsq-rco",
   "model": "Qwen3.8 Flash-Next IQ3_S (Strata)",
   "machine": "GMKtec NucBox EVO-X2",
   "hardware": "Ryzen AI Max+ 395, 128 GB, Radeon 8060S",
   "os": "Windows 11 Pro",
   "engine": "Strata 0.2.7 / ROCm 10.2",
   "context": "coding agent, 6 turns",
   "ttft_s": 4.5,
   "kind": "measured",
   "source": "https://github.com/Niko1221/Strata/blob/HEAD/docs/STRIX_HALO_WINDOWS.md",
   "note": "69% of the prompt came from the conversation cache; 4.5 s to first token per turn.",
   "decode_tok_s": 47.0,
   "prefill_tok_s": null
  },
  {
   "id": "strata-linux-8k",
   "model_id": "qwen3-8-flash-next-abliterated-gsq-rco-strata-gguf-gsq-rco",
   "model": "Qwen3.8 Flash-Next IQ3_S (Strata)",
   "machine": "Ryzen AI Max+ 395 test box",
   "hardware": "128 GB, Radeon 8060S, Linux",
   "os": "Linux",
   "engine": "Strata 0.2.x",
   "context": "8K",
   "ttft_s": null,
   "kind": "measured",
   "source": "https://github.com/Niko1221/Strata/blob/HEAD/docs/STRIX_HALO.md",
   "note": "Tuned switches. The same page measured 56.2 tok/s at 8K with defaults and 40.6-42.8 tok/s at 128K.",
   "decode_tok_s": 59.7,
   "prefill_tok_s": null
  },
  {
   "id": "maya-8060s",
   "model_id": null,
   "model": "GLM-5.3-Flash Maya-S",
   "machine": "Radeon 8060S test box",
   "hardware": "Ryzen AI Max+ 395, 128 GB, 96 GB for the GPU",
   "os": "Windows 11",
   "engine": "Project Maya 1.0.30",
   "context": "Maya-S",
   "ttft_s": null,
   "kind": "measured",
   "source": "https://github.com/mw00/project-maya",
   "note": "A user report for Maya-S, not the archived Maya-M. Short-context generation was 18.3 tok/s; prompt read 320 tok/s.",
   "decode_tok_s": 18.3,
   "prefill_tok_s": 320
  },
  {
   "id": "mirai-4070",
   "model_id": "qwen3-8-27b-s-mirai-gguf-mirai",
   "model": "Qwen3.8-27B-S Mirai",
   "machine": "RTX 4070",
   "hardware": "RTX 4070, 12 GB",
   "os": "Linux",
   "engine": "mirai-s-ada / CUDA 12",
   "context": "4K",
   "ttft_s": null,
   "kind": "measured",
   "source": "https://github.com/professorpalmer/mirai-s-ada",
   "note": "CUDA reference, not an M5 result. 60.9 tok/s at 64K, 20.3 at 128K, 10.3 at 200K; 1,090 tok/s prompt read.",
   "decode_tok_s": 70.1,
   "prefill_tok_s": 1090
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "strixfold-m5-firstlight",
   "model": "StrixFold \u00b7 DeepSeek V4.1 Flash 1017",
   "machine": "Bosgame M5 pair (third-party)",
   "hardware": "2\u00d7 Bosgame M5: Ryzen AI Max+ 395, 128 GB each, 2\u00d7 USB4",
   "os": "Linux, kernel 7.2.7, ROCm 10",
   "engine": "StrixFold HIP / EXL3 2.9 bpw",
   "context": "first light, one chat",
   "source": "https://github.com/Shoestring-dev/StrixFold/blob/HEAD/docs/BOSGAME-M5.md",
   "note": "Code 70.7 tok/s and prose 41.1 tok/s in the first-light test; one chat across the prompt set 58.2 tok/s; prompt read 1,284 tok/s. Stock 85 W balanced, drafting on.",
   "decode_tok_s": 58.2,
   "prefill_tok_s": 1284.0
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "strixfold-m5-agent-a",
   "model": "StrixFold \u00b7 DeepSeek V4.1 Flash 1017",
   "machine": "Bosgame M5 pair A (third-party)",
   "hardware": "2\u00d7 Bosgame M5: Ryzen AI Max+ 395, 128 GB each, USB4",
   "os": "Linux, kernel 7.2.7, ROCm 10",
   "engine": "StrixFold HIP / EXL3 2.9 bpw",
   "context": "agent work, <20K prompt",
   "source": "https://github.com/Shoestring-dev/StrixFold/blob/HEAD/docs/RESULTS.md",
   "note": "Median of 47 single requests; 20-150K prompts ranged 48.3-51.6 tok/s. 27.3 tok/s each with the prompt cache; stock balanced 85 W, up to four concurrent.",
   "decode_tok_s": 54.5,
   "prefill_tok_s": 1128.0
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "strixfold-m5-agent-b",
   "model": "StrixFold \u00b7 DeepSeek V4.1 Flash 1017",
   "machine": "Bosgame M5 pair B (third-party)",
   "hardware": "2\u00d7 Bosgame M5: Ryzen AI Max+ 395, 128 GB each, USB4",
   "os": "Linux, kernel 7.2.7, ROCm 10",
   "engine": "StrixFold HIP / EXL3 2.9 bpw",
   "context": "agent work, <20K prompt",
   "source": "https://github.com/Shoestring-dev/StrixFold/blob/HEAD/docs/RESULTS.md",
   "note": "Median of 37 single requests; 20-100K prompts ranged 39.7-52.1 tok/s; sample size is small and the USB4 links changed. 27.6 tok/s each with the prompt cache.",
   "decode_tok_s": 49.7,
   "prefill_tok_s": 1118.0
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "strixfold-halo",
   "model": "StrixFold \u00b7 DeepSeek V4.1 Flash 1017",
   "machine": "2\u00d7 Ryzen AI Halo (third-party)",
   "hardware": "2\u00d7 Ryzen AI Max+ 395, 128 GB each, 2\u00d7 USB4",
   "os": "Linux, kernel 7.2.7, ROCm 10",
   "engine": "StrixFold HIP / EXL3 2.9 bpw",
   "context": "one chat",
   "source": "https://github.com/Shoestring-dev/StrixFold/blob/HEAD/docs/RESULTS.md",
   "note": "1,543.4 tok/s prompt read at 11,594 tokens; eight chats total 107.96 tok/s; 72.3% of 2\u00d7 DGX Spark. Quality MMLU-Pro 84.8%, HumanEval+ 92%.",
   "decode_tok_s": 52.84,
   "prefill_tok_s": 1413.8
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "strixfold-halo-48k",
   "model": "StrixFold \u00b7 DeepSeek V4.1 Flash 1017",
   "machine": "2\u00d7 Ryzen AI Halo (third-party)",
   "hardware": "2\u00d7 Ryzen AI Max+ 395, 128 GB each, 2\u00d7 USB4",
   "os": "Linux, kernel 7.2.7, ROCm 10",
   "engine": "StrixFold HIP / EXL3 2.9 bpw",
   "context": "48K prompt, coding",
   "source": "https://github.com/Shoestring-dev/StrixFold/blob/HEAD/docs/RESULTS.md",
   "note": "Drafting on; 31.5 tok/s with drafting off. 32.7 at the task alone; 1M prompt read was 1,071 tok/s (962 s) and a cached second turn returned in 3.9 s.",
   "decode_tok_s": 73.7,
   "prefill_tok_s": 1379.0
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "tensorfold-qwen38-27b",
   "model": "TensorFold \u00b7 Qwen3.8-27B MLX-4bit",
   "machine": "Strix Halo (Radeon 8060S)",
   "hardware": "Radeon 8060S, gfx1151; board and RAM not named",
   "os": "Linux, ROCm",
   "engine": "TensorFold 1.0.12 HIP",
   "context": "code T0, community-v1",
   "source": "https://github.com/ashhart/TensorFold/issues/1001",
   "note": "Stream delivery tok/s. Chat T0 13.17, code T1 12.89, chat T1 12.90; first text 0.06-0.50 s. Local checkpoint receipts; not yet publishable on the board. Prefill is not reported in the receipt table.",
   "decode_tok_s": 12.88,
   "prefill_tok_s": null
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "tensorfold-qwen36-35b",
   "model": "TensorFold \u00b7 Qwen3.6-35B-A3B-MTP MLX-4bit",
   "machine": "Strix Halo (Radeon 8060S)",
   "hardware": "Radeon 8060S, gfx1151; board and RAM not named",
   "os": "Linux, ROCm",
   "engine": "TensorFold 1.0.12 HIP",
   "context": "code T0, community-v1",
   "source": "https://github.com/ashhart/TensorFold/issues/1001",
   "note": "Stream delivery tok/s. Chat T0 79.84, code T1 80.75, chat T1 72.96; first text 0.06-0.50 s. Local checkpoint receipts. Prefill is not reported in the receipt table.",
   "decode_tok_s": 103.36,
   "prefill_tok_s": null
  },
  {
   "model_id": null,
   "ttft_s": null,
   "kind": "measured",
   "id": "tensorfold-qwen35-9b",
   "model": "TensorFold \u00b7 Qwen3.5-9B MLX-4bit",
   "machine": "Strix Halo (Radeon 8060S)",
   "hardware": "Radeon 8060S, gfx1151; board and RAM not named",
   "os": "Linux, ROCm",
   "engine": "TensorFold 1.0.12 HIP",
   "context": "code T0, community-v1",
   "source": "https://github.com/ashhart/TensorFold/issues/1001",
   "note": "Stream delivery tok/s. Chat T0 38.95, code T1 38.61, chat T1 38.51; first text 0.06-0.50 s. Local checkpoint receipts. Prefill is not reported in the receipt table.",
   "decode_tok_s": 38.83,
   "prefill_tok_s": null
  }
 ],
 "pending": [
  {
   "machine": "Our Bosgame M5 x1",
   "reason": "No one-box result from our units yet. StrixFold measured pairs only; a single-box test is still needed."
  },
  {
   "machine": "Our Bosgame M5 x2, RPC",
   "reason": "StrixFold has third-party pair results below; our own pair, RPC split and load time are still pending."
  },
  {
   "machine": "MacBook Pro M4 Max / Mac mini M4",
   "reason": "No clean local result yet. The same test should run on Metal before the M5 arrives."
  },
  {
   "machine": "Full 262K context",
   "reason": "No measured result at the full Qwen3.8 window; the KV cache and speed both change with depth."
  },
  {
   "machine": "Image, video and TTS",
   "reason": "The page measures LLM token speed. Diffusion steps, video frames and real-time speech need their own benchmarks."
  }
 ]
}