Qwen 125B On Strix Halo Runs, But At What Speed?
Checked 2 October 2026. Benchmark results are the publisher’s own evaluations or one
person’s published measurements on their own machine; none is an independent reproduction.
The model: Qwen3.8 Flash Next
- Qwen3.8-Flash-Next model card:
“Number of Parameters: 125B with 6B activated, plus 51B n-gram embedding and 4B MTP”.
Mixture of experts: “Number of Experts: 512”, “Number of Activated Experts: 10 Routed + 1
Shared”. The card’s Safetensors panel lists the model size as 180B parameters, which is
125B + 51B + 4B. Context: 262,144 tokens natively, extensible to 1,000,000.
- Hosted version, same card: “Qwen3.8-Flash is the official version based on
Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default,
official built-in tools.”
-
Publisher coding table, same card (higher is better):
| Benchmark |
Flash Next |
Qwen3.8 27B |
DeepSeek V4 Flash 0731 |
| DeepSWE 1.1 |
58.7 |
42.2 |
54.4 |
| SWE-bench Pro |
62.5 |
61.7 |
56.0 |
| NL2Repo-Bench |
48.1 |
42.3 |
54.2 |
58.7 minus 42.2 is the 16.5 point DeepSWE lead. On NL2Repo-Bench the leader is DeepSeek V4
Flash 0731 at 54.2; Qwen3.8 27B scores 42.3, below Flash Next’s 48.1.
- On the Design of Qwen3.8-Next Architecture, submitted
31 August 2026: “a sparse mixture-of-experts model with 125B parameters, 6B activated per
token, and additional 51B parameters of n-gram embedding tables held off the
accelerator”, with “a single n-gram embedding layer whose tables are prefetched from host
memory”, and Qwen Sparse Attention scoring context with “a compressed lightweight indexer”.
Quantized downloads
- unsloth/Qwen3.8-Flash-Next-GGUF
lists the quantizations by size, including UD-IQ4_XS at 93.7 GB and UD-Q4_K_XL at 111 GB,
and separate MTP draft files (Q8_0 2.79 GB, BF16 7.77 GB).
- Four bit arithmetic: 125 billion parameters × 4 bits ÷ 8 bits per byte = 62.5 GB. This is a
floor on the core weights alone, not a file size.
The machine: AMD Strix Halo
- AMD Ryzen AI Halo developer platform:
AMD Ryzen AI Max+ 395 (16 cores, 32 threads, Zen 5); Radeon 8060S integrated graphics, 40
CUs, RDNA 3.5; memory LPDDR5x, 128GB, 8000MT/s; memory bandwidth 256GB/s; TDP 120W;
150 × 150 × 45.4 mm.
Measured speed on Strix Halo
- Qwen3.8-Flash-Next on Strix Halo, Vulkan only,
llama.cpp discussion #28512. A Bosgame M5 (Ryzen AI Max+ 395, Radeon 8060S, 128 GB) on
Vulkan with RADV, the stock UD-IQ4_XS quant and a Q8_0 draft head. Single stream, greedy.
Ten real agent conversations replayed: median decode 25 → 33 tokens per second after
tuning. Time to first token on a fresh 18K prompt 67 s → 44 s. Tuned decode at 8K: short
code 58.1, new code 41.6 (30.7 before tuning), prose 30.1, file rewrite 55.4 (63.2 with
n-max 6); at 32K new code 37.8, file rewrite 48.5 (55.2 with n-max 6). The biggest gain:
“Upstream sizes the expert-count shader for 256 experts, and this model has 512 … Lifting
the limit gave +19 % prefill.” Long drafts: “n-max 6 is the setting for file rewrites …
but it halves prose”.
- vulkan: raise the hoisted row-id limit for mul_mat_id from 256 to 512 experts,
llama.cpp pull request #28501, merged: prefill on Qwen3.8-Flash-Next Q5_K from 426 to 507
tokens per second for an 8K prompt (+19%), token generation unchanged.
- Qwen3.8-Flash-Next on Strix Halo (gfx1151): 17 → 47 tok/s,
llama.cpp discussion #27950, ROCm 7.1, UD-IQ4_XS, single stream, greedy. No speculation →
full stack: file rewrite at 8K 16.8 → 47.1; new code at 8K 16.8 → 31.7; file rewrite at 24K
about 15 → 28.6; new code at 24K about 15 → 25.4. The QSA indexer’s top-k “falls back to
CPU past ne=1024”; fixing it gave “+38–53% end-to-end at 24k”. Native MTP speculative
decoding.
Other hardware and hosted options
- NVIDIA GeForce RTX 5090:
32 GB of GDDR7. 93.7 GB of weights plus a 2.79 GB draft head is about 96.5 GB, which does
not fit in 32 GB without offloading.
- DeepSeek-V4-Flash-0731: NL2Repo
54.2 on its own card.
- DeepSeek-V4.1-Flash: a separate,
later release, a multimodal mixture of experts with a 552B backbone and contexts up to one
million tokens. Qwen’s table compares V4 Flash 0731, not V4.1.
Not checked, or not as stated
- The narration says Qwen3.8 27B leads NL2Repo-Bench 54.2 to 48.1, and later refers to “its
NL2Repo lead”. Qwen’s own card gives Qwen3.8 27B 42.3 on NL2Repo-Bench, below Flash Next’s
48.1; the 54.2 belongs to DeepSeek V4 Flash 0731.
- “30 for new code at 8K context” matches the Vulkan author’s pre-tuning figure (30.7); after
tuning the same row reads 41.6.
- How much of the 128 GB a given Strix Halo system lets the graphics side use depends on its
firmware and settings; no single figure is claimed.
- Benchmarks are the publisher’s own, and DeepSWE is reported as a model plus agent harness
score.