Local AI Registry

Recipe

qwen3-8-flash-next-unsloth-dynamic-iq4-xs-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1

qwen3-8-flash-next-unsloth-dynamic-iq4-xs-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1
JSON APIFind compatible recipes

Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
b10637 (strix-go-brrr / strix-halo-vulkan 39817c476)
Accelerators
1
Tensor parallel
1
Context tokens
32,768
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF
Repository
unsloth/Qwen3.8-Flash-Next-GGUF
Status
known
Link type
Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Observed configuration

llama.cpp

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmtbqotxt003gqq01kfluky16

Observed source tokens

Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.

  1. toolbox
  2. run
  3. -c
  4. flashnext-strix
  5. --
  6. llama-server
  7. -m
  8. ~/models/qwen3.8-flash-next/UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf
  9. -ngl
  10. 999
  11. --load-mode
  12. none
  13. -fa
  14. on
  15. -c
  16. 32768
  17. -ot
  18. per_layer_token_embd=CPU
  19. --spec-type
  20. ngram-mod
  21. --spec-ngram-mod-n-match
  22. 32
  23. --spec-draft-n-max
  24. 7
  25. -np
  26. 1
  27. --fit
  28. off
  29. -b
  30. 4096
  31. -ub
  32. 2048
  33. --temp
  34. 1.0
  35. --top-p
  36. 0.95
  37. --top-k
  38. 20
  39. --min-p
  40. 0.0
  41. --jinja
  42. --reasoning-format
  43. auto
  44. --host
  45. 0.0.0.0
  46. --port
  47. 8082
FlagValue
-cflashnext-strix
--llama-server
-m~/models/qwen3.8-flash-next/UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf
-ngl999
--load-modenone
-faon
-c32768
-otper_layer_token_embd=CPU
--spec-typengram-mod
--spec-ngram-mod-n-match32
--spec-draft-n-max7
-np1
--fitoff
-b4096
-ub2048
--temp1.0
--top-p0.95
--top-k20
--min-p0.0
--reasoning-formatauto
--host0.0.0.0
--port8082

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
132,768520.21103,267observedqwen3-8-flash-next-unsloth-dynamic-iq4-xs-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count
1
hardware id
ryzen-ai-max-plus-395-128gb
id
qwen3-8-flash-next-unsloth-dynamic-iq4-xs-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1
model instance id
unsloth-qwen3-8-flash-next-gguf--unsloth-dynamic-iq4-xs
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-8-flash-next-unsloth-dynamic-iq4-xs-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
b10637 (strix-go-brrr / strix-halo-vulkan 39817c476)

serving

max concurrency
1
max context tokens
32,768
tensor parallel
1
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
engine.graph mode · provenance · captured at
2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
serving.max concurrency · provenance · captured at
2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-derived-from-source-evidence
serving.max concurrency · state known

metadata

localmaxxing · batch size
1
localmaxxing · hardware label
Ryzen AI Max 395
localmaxxing · notes
strix-go-brrr engine (github.com/setianke/strix-go-brrr). Measurement: llama-server /v1/chat/completions API timings, code-rewrite task (1701 prompt tok, 800 output tok), warm cache run (1697 cached). Decode 110 t/s is ngram-mod speculation drafting from repeated code in context — cold novel-text baseline is 27 t/s unspeculated. Both numbers reported honestly; the workload is representative of agentic coding, not cherry-picked. Container: Vulkan/RADV, flashnext-strix toolbox from strix-halo-vulkan fork (commit 39817c476, qwen4exp arch). N-gram embedding table (51B params) offloaded to CPU via -ot per_layer_token_embd=CPU — zero speed cost, frees ~27 GB GTT. KV f16 mandatory (quantised KV crashes on qwen4exp dense-attention layers). Day-1 build. No MTP drafter (not in Unsloth GGUFs), no EVICT adaptive verification, no ubatch tuning, no profiling. Room to grow.
localmaxxing · observed command
toolbox run -c flashnext-strix -- llama-server -m ~/models/qwen3.8-flash-next/UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf -ngl 999 --load-mode none -fa on -c 32768 -ot per_layer_token_embd=CPU --spec-type ngram-mod --spec-ngram-mod-n-match 32 --spec-draft-n-max 7 -np 1 --fit off -b 4096 -ub 2048 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --jinja --reasoning-format auto --host 0.0.0.0 --port 8082
localmaxxing · run id
cmtbqotxt003gqq01kfluky16
localmaxxing · tokenized · arguments
toolbox, run, -c, flashnext-strix, --, llama-server, -m, ~/models/qwen3.8-flash-next/UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf, -ngl, 999, --load-mode, none, -fa, on, -c, 32768, -ot, per_layer_token_embd=CPU, --spec-type, ngram-mod, --spec-ngram-mod-n-match, 32, --spec-draft-n-max, 7, -np, 1, --fit, off, -b, 4096, -ub, 2048, --temp, 1.0, --top-p, 0.95, --top-k, 20, --min-p, 0.0, --jinja, --reasoning-format, auto, --host, 0.0.0.0, --port, 8082
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:10:02Z

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmtbqotxt003gqq01kfluky16 ↗