shared configs · hal0-profiles

profiles

A profile is the flag string that produced a number, versioned and downloadable. Find the one for your model, install it with the CLI, or read the config and steal the parts you want.

how profiles work submit a profile

model

intent

lane

8 of 8 profiles

profile · v3 · rocmagent

agent-27b-tools

Tool-calling agent slot

runs these models

chadrockchadrock3.6-27b-pi-agentchadrockchadrock3-6-27b-pi-agent

35.5tok/s

decode tok/s

301tok/s

prefill

measured on chadrock3.6-27b-pi-agent-rocmfp4-mtp · rocm

@jcbtcconfig & flags →

profile · v2 · rocmcoding

coder-next-40b

40B MoE coder, editor-attached

runs these models

qwen3-coderqwen3-coder-next-reap-40b-a3bqwen3-coderqwen3-coder-reap-25b-a3b

54.7tok/s

decode tok/s

1368tok/s

prefill

measured on qwen3-coder-reap-25b-a3b-q5km · rocm

@lovedheartconfig & flags →

profile · v3 · rocmmoe

kv-q8-longctx

16k context on a 35B MoE

runs these models

qwen3qwen3.6-35b-a3bgemmagemma-4-26b-a4b

46.1tok/s

decode tok/s

1300tok/s

prefill

measured on qwen3.6-35b-a3b-q4kxl · rocm

@kyuz0config & flags →

profile · v6 · rocmchat

lean-9b-daily

The everyday 9B

runs these models

qwen3qwen3.5-9bqwen3qwen3.5-9b-deepseek-v4-flash

51.3tok/s

decode tok/s

613tok/s

prefill

measured on qwen3.5-9b-deepseek-v4-flash-mtp · rocm

@kyuz0config & flags →

profile · v4 · rocmchat

strix-mtp-max

Speculative decode, everything on

runs these models

chadrockchadrock3-6-35b-uncensoredchadrockchadrock-35b-ace-saber+1

102.1tok/s

decode tok/s

890tok/s

prefill

measured on chadrock3-6-35b-uncensored-mtp-strix-lean · rocm

first-partyconfig & flags →

profile · v1 · rocmdraft

tiny-draft-0-8b

Draft model for speculative pairs

runs these models

qwen3qwen3.5-0.8bqwen3qwen3-zero-coder-v2-0.8b

169.8tok/s

decode tok/s

6248tok/s

prefill

measured on qwen3.5-0.8b · rocm

first-partyconfig & flags →

profile · v2 · rocmvision

vision-27b

Screenshots and PDFs

runs these models

qwen3qwen3.6-27b-heretic

11.6tok/s

decode tok/s

295tok/s

prefill

measured on qwen3.6-27b-heretic-q4km · rocm

@nx-voidconfig & flags →

profile · v2 · vulkan_radvchat

vulkan-fallback

When ROCm will not build

runs these models

qwen3qwen3.6-35b-a3bgemmagemma-4-26b-a4b

46.1tok/s

decode tok/s

1300tok/s

prefill

measured on qwen3.6-35b-a3b-q4kxl · vulkan_radv

@lhlconfig & flags →

your profile here

<your-slug>

Found a flag string that makes your box sing? Version it, PR it, and it ships here with the numbers it produced.

submit a profile →

profile · v3 · 2026-06-17

agent-27b-tools

Tool-calling agent slot

q8 KV cache because tool traces are long and lossy cache breaks JSON. 79.8 % accept on draft-mtp.

install with hal0

install

hal0 profile install agent-27b-tools

flags

llama-server

-ngl 99 -c 8192 -fa 1 -ctk q8_0 -ctv q8_0 --parallel 1 -b 512 --draft-max 4 --mtp on

config

agent-27b-tools.toml

            schema = 1
 
[profile]
slug = "agent-27b-tools"
title = "Tool-calling agent slot"
summary = "q8 KV cache because tool traces are long and lossy cache breaks JSON. 79.8 % accept on draft-mtp."
intent = "agent"
author = "jcbtc"
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "chadrock3.6-27b-pi-agent-rocmfp4-mtp"
compatible = ["chadrock3-6-27b-pi-agent-mtp-rocmfp4-strix-lean"]
 
[args]
raw = "-ngl 99 -c 8192 -fa 1 -ctk q8_0 -ctv q8_0 --parallel 1 -b 512 --draft-max 4 --mtp on"
 
[[history]]
v = 3
date = "2026-06-17"
note = "q8 KV cache — tool traces stopped producing malformed JSON"
 
[[history]]
v = 2
date = "2026-05-19"
note = "-c 8192 for longer tool transcripts"
 
[[history]]
v = 1
date = "2026-05-02"
note = "initial"

runs produced by this config

rocm

model	decode	prefill	gb
chadrock3.6-27b-pi-agent-rocmfp4-mtp	35.5	301	14.8
chadrock3-6-27b-pi-agent-mtp-rocmfp4-strix-lean	34.9	308	14.8

versions

v32026-06-17q8 KV cache — tool traces stopped producing malformed JSON

v22026-05-19-c 8192 for longer tool transcripts

v12026-05-02initial

profile · v2 · 2026-06-14

coder-next-40b

40B MoE coder, editor-attached

Sized for autocomplete rather than long generations: 28.5 GB resident, prefill above 750 tok/s so a whole file lands fast.

install with hal0

install

hal0 profile install coder-next-40b

flags

llama-server

-ngl 99 -c 8192 -fa 1 -ctk q4_0 -ctv q4_0 --parallel 2 -b 1024

config

coder-next-40b.toml

            schema = 1
 
[profile]
slug = "coder-next-40b"
title = "40B MoE coder, editor-attached"
summary = "Sized for autocomplete rather than long generations: 28.5 GB resident, prefill above 750 tok/s so a whole file lands fast."
intent = "coding"
author = "lovedheart"
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "qwen3-coder-next-reap-40b-a3b-q4kxl"
compatible = ["qwen3-coder-reap-25b-a3b-q5km"]
 
[args]
raw = "-ngl 99 -c 8192 -fa 1 -ctk q4_0 -ctv q4_0 --parallel 2 -b 1024"
 
[[history]]
v = 2
date = "2026-06-14"
note = "-b 1024 and --parallel 2 — prefill +6 %, editor feels immediate"
 
[[history]]
v = 1
date = "2026-05-09"
note = "initial"

runs produced by this config

rocm

model	decode	prefill	gb
qwen3-coder-reap-25b-a3b-q5km	54.7	1368	17.7
qwen3-coder-next-reap-40b-a3b-q4kxl	26.8	756	28.5

versions

v22026-06-14-b 1024 and --parallel 2 — prefill +6 %, editor feels immediate

v12026-05-09initial

profile · v3 · 2026-06-19

kv-q8-longctx

16k context on a 35B MoE

q8 KV cache trades 3 % decode for context that actually holds a repo map. 22.4 GB resident at 16k.

install with hal0

install

hal0 profile install kv-q8-longctx

flags

llama-server

-ngl 99 -c 16384 -fa 1 -ctk q8_0 -ctv q8_0 --parallel 1 -b 1024

config

kv-q8-longctx.toml

            schema = 1
 
[profile]
slug = "kv-q8-longctx"
title = "16k context on a 35B MoE"
summary = "q8 KV cache trades 3 % decode for context that actually holds a repo map. 22.4 GB resident at 16k."
intent = "moe"
author = "kyuz0"
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "qwen3.6-35b-a3b-q4kxl"
compatible = ["gemma-4-26b-a4b-it-q4kxl"]
 
[args]
raw = "-ngl 99 -c 16384 -fa 1 -ctk q8_0 -ctv q8_0 --parallel 1 -b 1024"
 
[[history]]
v = 3
date = "2026-06-19"
note = "-c 16384 at q8 cache — 22.4 GB resident, 3 % decode given up"
 
[[history]]
v = 2
date = "2026-05-31"
note = "-b 1024"
 
[[history]]
v = 1
date = "2026-05-10"
note = "initial"

runs produced by this config

rocm

model	decode	prefill	gb
qwen3.6-35b-a3b-q4kxl	46.1	1300	22.4
gemma-4-26b-a4b-it-q4kxl	40.9	1336	17.1

versions

v32026-06-19-c 16384 at q8 cache — 22.4 GB resident, 3 % decode given up

v22026-05-31-b 1024

v12026-05-10initial

profile · v6 · 2026-06-18

lean-9b-daily

The everyday 9B

6 GB resident, leaves the pool free for embed, stt and image gen. What most boxes should run on the primary slot.

install with hal0

install

hal0 profile install lean-9b-daily

flags

llama-server

-ngl 99 -c 16384 -fa 1 -ctk q4_0 -ctv q4_0 --parallel 1 -b 512

config

lean-9b-daily.toml

            schema = 1
 
[profile]
slug = "lean-9b-daily"
title = "The everyday 9B"
summary = "6 GB resident, leaves the pool free for embed, stt and image gen. What most boxes should run on the primary slot."
intent = "chat"
author = "kyuz0"
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "qwen3.5-9b-q4kxl"
compatible = ["qwen3.5-9b-deepseek-v4-flash-mtp"]
 
[args]
raw = "-ngl 99 -c 16384 -fa 1 -ctk q4_0 -ctv q4_0 --parallel 1 -b 512"
 
[[history]]
v = 6
date = "2026-06-18"
note = "-c 16384 — fits alongside embed and stt on 96 GB GTT"
 
[[history]]
v = 5
date = "2026-05-30"
note = "drop --mlock; ZFS ARC was fighting it"
 
[[history]]
v = 4
date = "2026-05-12"
note = "q4_0 KV cache, 1.4 GB back"
 
[[history]]
v = 3
date = "2026-04-27"
note = "-fa 1"
 
[[history]]
v = 2
date = "2026-04-15"
note = "raise -ngl to 99"
 
[[history]]
v = 1
date = "2026-03-28"
note = "initial"

runs produced by this config

rocm

model	decode	prefill	gb
qwen3.5-9b-deepseek-v4-flash-mtp	51.3	613	7.6
qwen3.5-9b-q4kxl	33.0	1045	6

versions

v62026-06-18-c 16384 — fits alongside embed and stt on 96 GB GTT

v52026-05-30drop --mlock; ZFS ARC was fighting it

v42026-05-12q4_0 KV cache, 1.4 GB back

v32026-04-27-fa 1

v22026-04-15raise -ngl to 99

v12026-03-28initial

profile · v4 · 2026-06-19

strix-mtp-max

Speculative decode, everything on

draft-mtp with an f16 KV cache. The fastest tokens this box produces, at the cost of 19 GB resident and no room for a second big slot.

install with hal0

install

hal0 profile install strix-mtp-max

flags

llama-server

-ngl 99 -c 2048 -fa 1 -ctk f16 -ctv f16 --parallel 1 -b 512 --draft-max 4 --draft-min 1 --mtp on

config

strix-mtp-max.toml

            schema = 1
 
[profile]
slug = "strix-mtp-max"
title = "Speculative decode, everything on"
summary = "draft-mtp with an f16 KV cache. The fastest tokens this box produces, at the cost of 19 GB resident and no room for a second big slot."
intent = "chat"
author = "hal0-ci"
first_party = true
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "chadrock3-6-35b-uncensored-mtp-strix-lean"
compatible = ["chadrock-35b-ace-saber", "qwen3.6-35b-a3b-crown-halo-mtp-dynamic"]
 
[args]
raw = "-ngl 99 -c 2048 -fa 1 -ctk f16 -ctv f16 --parallel 1 -b 512 --draft-max 4 --draft-min 1 --mtp on"
 
[[history]]
v = 4
date = "2026-06-19"
note = "pin build b9219 — draft-mtp accept regressed on b9101"
 
[[history]]
v = 3
date = "2026-05-22"
note = "--draft-max 4 (was 8); shorter drafts accept more often here"
 
[[history]]
v = 2
date = "2026-04-30"
note = "f16 KV cache — q4 cache cost 9 % accept on this pair"
 
[[history]]
v = 1
date = "2026-04-11"
note = "initial"

runs produced by this config

rocm

model	decode	prefill	gb
chadrock3-6-35b-uncensored-mtp-strix-lean	102.1	890	19
chadrock-35b-ace-saber	100.5	903	19
qwen3.6-35b-a3b-crown-halo-mtp-dynamic	84.4	873	22.6

versions

v42026-06-19pin build b9219 — draft-mtp accept regressed on b9101

v32026-05-22--draft-max 4 (was 8); shorter drafts accept more often here

v22026-04-30f16 KV cache — q4 cache cost 9 % accept on this pair

v12026-04-11initial

profile · v1 · 2026-06-19

tiny-draft-0-8b

Draft model for speculative pairs

0.6 GB. Not a chat model — the draft half of a speculative pair, pinned so the big model never waits on it.

install with hal0

install

hal0 profile install tiny-draft-0-8b

flags

llama-server

-ngl 99 -c 2048 -fa 1 -ctk q4_0 --parallel 1 -b 256 --draft-p-min 0.6

config

tiny-draft-0-8b.toml

            schema = 1
 
[profile]
slug = "tiny-draft-0-8b"
title = "Draft model for speculative pairs"
summary = "0.6 GB. Not a chat model — the draft half of a speculative pair, pinned so the big model never waits on it."
intent = "draft"
author = "hal0-ci"
first_party = true
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "qwen3.5-0.8b"
compatible = ["qwen3-zero-coder-v2-0.8b-f16"]
 
[args]
raw = "-ngl 99 -c 2048 -fa 1 -ctk q4_0 --parallel 1 -b 256 --draft-p-min 0.6"
 
[[history]]
v = 1
date = "2026-06-19"
note = "initial — shipped with the june sweep"

runs produced by this config

rocm

model	decode	prefill	gb
qwen3.5-0.8b	169.8	6248	0.6
qwen3-zero-coder-v2-0.8b-f16	76.3	4827	1.6

versions

v12026-06-19initial — shipped with the june sweep

profile · v2 · 2026-06-11

vision-27b

Screenshots and PDFs

Vision projector loaded, 16.9 GB resident. Decode is slow — this slot reads, it does not chat.

install with hal0

install

hal0 profile install vision-27b

flags

llama-server

-ngl 99 -c 4096 -fa 1 -ctk q4_0 -ctv q4_0 --mmproj mmproj-f16.gguf --parallel 1

config

vision-27b.toml

            schema = 1
 
[profile]
slug = "vision-27b"
title = "Screenshots and PDFs"
summary = "Vision projector loaded, 16.9 GB resident. Decode is slow — this slot reads, it does not chat."
intent = "vision"
author = "nx-void"
 
[runner]
kind = "llama-server"
lane = "rocm"
min_build = "b9219"
image = "ghcr.io/hal0ai/amd-strix-halo-toolboxes:rocm-7.2.4-rocmfp4-server"
 
[model]
id = "qwen3.6-27b-heretic-q4km"
 
[args]
raw = "-ngl 99 -c 4096 -fa 1 -ctk q4_0 -ctv q4_0 --mmproj mmproj-f16.gguf --parallel 1"
 
[[history]]
v = 2
date = "2026-06-11"
note = "mmproj at f16 — q8 projector blurred small text in screenshots"
 
[[history]]
v = 1
date = "2026-05-24"
note = "initial"

runs produced by this config

rocm

model	decode	prefill	gb
qwen3.6-27b-heretic-q4km	11.6	295	16.9

versions

v22026-06-11mmproj at f16 — q8 projector blurred small text in screenshots

v12026-05-24initial

profile · v2 · 2026-06-16

vulkan-fallback

When ROCm will not build

vulkan_radv lane, same models. Roughly 18 % slower decode and 29 % slower prefill, but it starts on a stock kernel.

install with hal0

install

hal0 profile install vulkan-fallback

flags

llama-server

-ngl 99 -c 2048 -fa 1 -ctk q4_0 -ctv q4_0 --parallel 1 -b 512 --device Vulkan0

config

vulkan-fallback.toml

            schema = 1
 
[profile]
slug = "vulkan-fallback"
title = "When ROCm will not build"
summary = "vulkan_radv lane, same models. Roughly 18 % slower decode and 29 % slower prefill, but it starts on a stock kernel."
intent = "chat"
author = "lhl"
 
[runner]
kind = "llama-server"
lane = "vulkan_radv"
 
[model]
id = "qwen3.6-35b-a3b-q4kxl"
compatible = ["gemma-4-26b-a4b-it-q4kxl"]
 
[args]
raw = "-ngl 99 -c 2048 -fa 1 -ctk q4_0 -ctv q4_0 --parallel 1 -b 512 --device Vulkan0"
 
[[history]]
v = 2
date = "2026-06-16"
note = "--device Vulkan0 — stops radv picking the wrong node in an LXC"
 
[[history]]
v = 1
date = "2026-06-02"
note = "initial"

runs produced by this config

vulkan_radv

model	decode	prefill	gb
qwen3.6-35b-a3b-q4kxl	46.1	1300	22.4
gemma-4-26b-a4b-it-q4kxl	40.9	1336	17.1

versions

v22026-06-16--device Vulkan0 — stops radv picking the wrong node in an LXC

v12026-06-02initial