Recipe
qwen3-6-35b-a3b-q4-k-m-intel-arc-pro-b70-32gb-llama-cpp-tp1
qwen3-6-35b-a3b-q4-k-m-intel-arc-pro-b70-32gb-llama-cpp-tp1Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- llama.cpp
- Engine version
- upstream llama.cpp b9553 (Vulkan backend)
- Accelerators
- 1
- Tensor parallel
- 1
- Context tokens
- 262,144
- Max concurrency
- 2,048
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-v1-MTP-GGUF- Repository
- Jackrong/Qwopus3.6-35B-A3B-v1-MTP-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
Observed source tokens
Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.
llama-server-mQwopus3.6-35B-A3B-v1-MTP-Q4_K_M.gguf--host0.0.0.0--port8080-ngl99-c262144-np1--deviceVulkan2--jinja--no-webui--reasoning-formatdeepseek
| Flag | Value |
|---|---|
-m | Qwopus3.6-35B-A3B-v1-MTP-Q4_K_M.gguf |
--host | 0.0.0.0 |
--port | 8080 |
-ngl | 99 |
-c | 262144 |
-np | 1 |
--device | Vulkan2 |
--reasoning-format | deepseek |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 2048 | 262,144 | — | 107.8 | — | observed | qwen3-6-35b-a3b-q4-k-m-intel-arc-pro-b70-32gb-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
- intel-arc-pro-b70-32gb
- id
- qwen3-6-35b-a3b-q4-k-m-intel-arc-pro-b70-32gb-llama-cpp-tp1
- model instance id
- jackrong-qwopus3-6-35b-a3b-v1-mtp-gguf--q4-k-m
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- qwen3-6-35b-a3b-q4-k-m-intel-arc-pro-b70-32gb-llama-cpp-tp1-sweep
- status
- candidate
capabilities
engine
- name
- llama.cpp
- version
- upstream llama.cpp b9553 (Vulkan backend)
serving
- max concurrency
- 2,048
- max context tokens
- 262,144
- 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
metadata
- localmaxxing · backend
- vulkan
- localmaxxing · hardware label
- Intel Arc Pro B70
- localmaxxing · notes
- Upstream llama.cpp b9553 (Vulkan backend), running natively on Windows 11 (NOT through WSL — Vulkan B70 path requires native Windows). Intel Arc Pro B70 32GB. Critical flag: --device Vulkan2 is REQUIRED to pin execution to the B70. Without it, llama.cpp layer-splits across detected Vulkan devices (RTX 3090 + Intel iGPU + B70), and the iGPU running on system RAM becomes the bottleneck — observed regression from 107 t/s to ~22 t/s. Single-device pinning gives stable performance. Known regressions on this build/model combo (DO NOT USE): - -fa / flash attention: crashes or severe degradation - -ctk q8_0 -ctv q8_0: 107 -> 8-13 t/s (Vulkan KV-quant path unoptimized) - --spec-type draft-mtp: 107 -> 27 t/s (4x regression — confirmed by upstream issue ggml-org/llama.cpp#23769, Qwen3.6-35B-A3B MoE class on Intel Arc Vulkan crashes/degrades; PR #24312 attempted fix was closed, bug remains unfixed in upstream master as of build b9553) Bench methodology: 5 sequential completions of "Write a 300 word explanation of the rule against perpetuities" at temperature 0.7, max_tokens=400, 26-token prompt, via /v1/chat/completions. Token/sec values from llama.cpp predicted_per_second timing field. Individual runs (t/s): paste your actual 5 numbers here Median: 107.78 t/s, mean ~107.5 t/s. System: AMD Ryzen 7 9800X3D, 128GB DDR5, Windows 11 native (B70 driver only available on Windows, not WSL2).
- localmaxxing · observed command
- llama-server \ -m Qwopus3.6-35B-A3B-v1-MTP-Q4_K_M.gguf \ --host 0.0.0.0 --port 8080 \ -ngl 99 -c 262144 -np 1 \ --device Vulkan2 \ --jinja --no-webui \ --reasoning-format deepseek
- localmaxxing · run id
- cmqd2eqmk00e8mp01u1fq6h3u
- localmaxxing · tokenized · arguments
- llama-server, -m, Qwopus3.6-35B-A3B-v1-MTP-Q4_K_M.gguf, --host, 0.0.0.0, --port, 8080, -ngl, 99, -c, 262144, -np, 1, --device, Vulkan2, --jinja, --no-webui, --reasoning-format, deepseek
- localmaxxing · tokenized · fidelity
- faithful
provenance
- captured at
- 2026-08-30T09:10:02Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-30T09:10:02Z | normalized-recipe | www.localmaxxing.com/en/runs/cmqd2eqmk00e8mp01u1fq6h3u ↗ |