Bench Board · local model benchmark for opencode

Generated 14/10/2026 11:28 · scores computed with the harness functions (same math as the CLI)
30 models1800 tasksweights tools 0.30 · code 0.30 · agent 0.25 · speed 0.15NVIDIA GeForce RTX 5080measured 2026-10-10 → 2026-10-14head-to-head compare →gray = retired model (record only)task sets 16/60/60 · 12/60/28 · 8/60/24engines llama unknown · llama 0.6.0-dev+4413a19 · llama 0.4.0-dev · strata 0.1.43.1 · strata 0.1.44.1
0.84
Best overall
307 t/s
Fastest (gen)
13 ms
Lowest latency
2026-10-14
Newest model

Weighted global

static order: global — the control reorders in place, no network
1
gemma-4-26b-a4b-it-ud-q5_k_xl2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.84 ±1 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off
2
qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 0
0.84 v3·8/60/24 llama 0.4.0-dev think:on retired
3
qwen3_8-27b-ud-q5_k_m2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 4
0.83 v3·8/60/24 llama unknown think:on retired
4
qwen3_8-27b-ud-q3_k_xl2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 2
0.83 v3·8/60/24 llama unknown think:on
5
qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 1
0.83 v3·8/60/24 llama unknown think:on
6
qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.81 ±1 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off
7
qwen36-35b-distill-mxfp4_moe2026-10-14 · 16+60+60 tasks · version v3·16/60/60 · excluded code: 0
0.81 ±0 v3·16/60/60 llama 0.6.0-dev+4413a19 think:off
8
ornith-1_5-35b-a3b-apex-mtp-quality2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 1
0.81 v3·8/60/24 llama unknown think:on
9
huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 0
0.80 v3·8/60/24 llama unknown think:on retired
10
strata-iq3_xxs2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.80 ±2 v3·12/60/28 strata 0.1.43.1 think:off ★ optimal
11
qwen3_8-27b-heretic-q4_k_m2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.79 ±3 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off
12
qwen3-vl-8b-instruct-1m-q6_k2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 0
0.79 v3·8/60/24 llama unknown think:on
13
ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 0
0.79 v3·8/60/24 llama unknown think:on retired
14
qwen3_8-27b-ud-q3_k_xl2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.78 ±2 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off
15
qwen36-35b-distill-mxfp4_moe2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 5
0.78 v3·8/60/24 llama unknown think:on
16
ornith-1_5-35b-a3b-apex-mtp-quality2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.78 ±2 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off
17
qwen3_8-27b-heretic-q4_k_m2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 1
0.78 v3·8/60/24 llama unknown think:on
18
qwen3_8-27b-ud-q5_k_m2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.77 ±3 v3·12/60/28 llama 0.4.0-dev think:off retired
19
qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp2026-10-12 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.77 ±3 v3·12/60/28 llama 0.4.0-dev think:off retired
20
qwen3-vl-8b-instruct-1m-q6_k2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.77 ±2 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off
21
huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.76 ±1 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off retired
22
qwen3_8-27b-ud-q4_k_m2026-10-12 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.76 ±2 v3·12/60/28 llama 0.4.0-dev think:off retired
23
strata-iq3_xxs2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.76 ±0 v3·12/60/28 strata 0.1.43.1 think:on
24
ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs2026-10-13 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.76 ±5 v3·12/60/28 llama 0.6.0-dev+4413a19 think:off retired
25
qwen3_8-27b-ud-q4_k_xl2026-10-12 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.75 ±2 v3·12/60/28 llama 0.4.0-dev think:off retired
26
nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs2026-10-12 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 6
0.75 ±4 v3·12/60/28 llama 0.4.0-dev think:off retired
27
strata-swift-iq3_xxs2026-10-14 · 16+60+60 tasks · version v3·16/60/60 · excluded code: 0
0.73 ±4 v3·16/60/60 strata 0.1.44.1 think:off
28
gemma-4-26b-a4b-it-ud-q5_k_xl2026-10-10 · 8+60+24 tasks · version v3·8/60/24 · excluded code: 0
0.71 v3·8/60/24 llama unknown think:on
29
muse-glimmer-30b-kquant-dynamic-q4_k_xl2026-10-12 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 0
0.58 ±4 v3·12/60/28 llama 0.4.0-dev think:off retired
30
qwen3-coder-30b-a3b-instruct-ud-q4_k_xl2026-10-12 · 12+60+28 tasks · version v3·12/60/28 · excluded code: 29
0.45 ±7 v3·12/60/28 llama 0.4.0-dev think:off retired

By task set

A score is comparable only inside a set: the pool may hold rows measured with different task counts (the SET badge). See docs/MILLORA.md §C9.

16/60/60 tasks · label v3 · 2 rows · measured 2026-10-14

qwen36-35b-distill-mxfp4_moe2026-10-14 · tools 94% · code 76% · agent 77%
0.81 think:offv3·16/60/60
strata-swift-iq3_xxs2026-10-14 · tools 84% · code 77% · agent 78%
0.73 think:offv3·16/60/60

12/60/28 tasks · label v3 · 17 rows · measured 2026-10-12 → 2026-10-13

gemma-4-26b-a4b-it-ud-q5_k_xl2026-10-13 · tools 92% · code 91% · agent 93%
0.84 think:offv3·12/60/28
qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s2026-10-13 · tools 92% · code 84% · agent 96%
0.81 think:offv3·12/60/28
strata-iq3_xxs2026-10-13 · tools 86% · code 88% · agent 96%
0.80 think:offv3·12/60/28
qwen3_8-27b-heretic-q4_k_m2026-10-13 · tools 92% · code 87% · agent 92%
0.79 think:offv3·12/60/28
qwen3_8-27b-ud-q3_k_xl2026-10-13 · tools 92% · code 76% · agent 92%
0.78 think:offv3·12/60/28
ornith-1_5-35b-a3b-apex-mtp-quality2026-10-13 · tools 92% · code 73% · agent 87%
0.78 think:offv3·12/60/28
qwen3_8-27b-ud-q5_k_m2026-10-13 · tools 92% · code 78% · agent 87%
0.77 think:off retiredv3·12/60/28
qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp2026-10-12 · tools 92% · code 76% · agent 92%
0.77 think:off retiredv3·12/60/28
qwen3-vl-8b-instruct-1m-q6_k2026-10-13 · tools 92% · code 76% · agent 87%
0.77 think:offv3·12/60/28
huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s2026-10-13 · tools 92% · code 72% · agent 82%
0.76 think:off retiredv3·12/60/28
qwen3_8-27b-ud-q4_k_m2026-10-12 · tools 92% · code 68% · agent 93%
0.76 think:off retiredv3·12/60/28
strata-iq3_xxs2026-10-13 · tools 83% · code 89% · agent 96%
0.76 think:onv3·12/60/28
ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs2026-10-13 · tools 92% · code 75% · agent 77%
0.76 think:off retiredv3·12/60/28
qwen3_8-27b-ud-q4_k_xl2026-10-12 · tools 92% · code 72% · agent 82%
0.75 think:off retiredv3·12/60/28
nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs2026-10-12 · tools 83% · code 44% · agent 87%
0.75 think:off retiredv3·12/60/28
muse-glimmer-30b-kquant-dynamic-q4_k_xl2026-10-12 · tools 33% · code 81% · agent 93%
0.58 think:off retiredv3·12/60/28
qwen3-coder-30b-a3b-instruct-ud-q4_k_xl2026-10-12 · tools 92% · code 23% · agent 0%
0.45 think:off retiredv3·12/60/28

8/60/24 tasks · label v3 · 11 rows · measured 2026-10-10

qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp2026-10-10 · tools 100% · code 95% · agent 100%
0.84 think:on retiredv3·8/60/24
qwen3_8-27b-ud-q5_k_m2026-10-10 · tools 100% · code 96% · agent 96%
0.83 think:on retiredv3·8/60/24
qwen3_8-27b-ud-q3_k_xl2026-10-10 · tools 100% · code 95% · agent 96%
0.83 think:onv3·8/60/24
qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s2026-10-10 · tools 100% · code 92% · agent 100%
0.83 think:onv3·8/60/24
ornith-1_5-35b-a3b-apex-mtp-quality2026-10-10 · tools 100% · code 92% · agent 92%
0.81 think:onv3·8/60/24
huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s2026-10-10 · tools 100% · code 90% · agent 92%
0.80 think:on retiredv3·8/60/24
qwen3-vl-8b-instruct-1m-q6_k2026-10-10 · tools 100% · code 73% · agent 88%
0.79 think:onv3·8/60/24
ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs2026-10-10 · tools 88% · code 93% · agent 96%
0.79 think:on retiredv3·8/60/24
qwen36-35b-distill-mxfp4_moe2026-10-10 · tools 100% · code 78% · agent 96%
0.78 think:onv3·8/60/24
qwen3_8-27b-heretic-q4_k_m2026-10-10 · tools 100% · code 90% · agent 83%
0.78 think:onv3·8/60/24
gemma-4-26b-a4b-it-ud-q5_k_xl2026-10-10 · tools 75% · code 87% · agent 88%
0.71 think:onv3·8/60/24

Code (60 tasks · 4 languages)

think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl
23%
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs
44%
think:off retired qwen3_8-27b-ud-q4_k_m
68%
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
72%
think:off retired qwen3_8-27b-ud-q4_k_xl
72%
think:on qwen3-vl-8b-instruct-1m-q6_k
73%
think:off ornith-1_5-35b-a3b-apex-mtp-quality
73%
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
75%
think:off qwen36-35b-distill-mxfp4_moe
76%
think:off qwen3_8-27b-ud-q3_k_xl
76%
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
76%
think:off qwen3-vl-8b-instruct-1m-q6_k
76%
think:off strata-swift-iq3_xxs
77%
think:on qwen36-35b-distill-mxfp4_moe
78%
think:off retired qwen3_8-27b-ud-q5_k_m
78%
think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl
81%
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
84%
think:off qwen3_8-27b-heretic-q4_k_m
87%
think:on gemma-4-26b-a4b-it-ud-q5_k_xl
87%
think:off strata-iq3_xxs
88%
think:on strata-iq3_xxs
89%
think:on qwen3_8-27b-heretic-q4_k_m
90%
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
90%
think:off gemma-4-26b-a4b-it-ud-q5_k_xl
91%
think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
92%
think:on ornith-1_5-35b-a3b-apex-mtp-quality
92%
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
93%
think:on qwen3_8-27b-ud-q3_k_xl
95%
think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
95%
think:on retired qwen3_8-27b-ud-q5_k_m
96%

Multi-step agent (60 tasks)

think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl
0%
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
77%
think:off qwen36-35b-distill-mxfp4_moe
77%
think:off strata-swift-iq3_xxs
78%
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
82%
think:off retired qwen3_8-27b-ud-q4_k_xl
82%
think:on qwen3_8-27b-heretic-q4_k_m
83%
think:off ornith-1_5-35b-a3b-apex-mtp-quality
87%
think:off retired qwen3_8-27b-ud-q5_k_m
87%
think:off qwen3-vl-8b-instruct-1m-q6_k
87%
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs
87%
think:on qwen3-vl-8b-instruct-1m-q6_k
88%
think:on gemma-4-26b-a4b-it-ud-q5_k_xl
88%
think:on ornith-1_5-35b-a3b-apex-mtp-quality
92%
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
92%
think:off qwen3_8-27b-heretic-q4_k_m
92%
think:off qwen3_8-27b-ud-q3_k_xl
92%
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
92%
think:off gemma-4-26b-a4b-it-ud-q5_k_xl
93%
think:off retired qwen3_8-27b-ud-q4_k_m
93%
think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl
93%
think:on retired qwen3_8-27b-ud-q5_k_m
96%
think:on qwen3_8-27b-ud-q3_k_xl
96%
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
96%
think:on qwen36-35b-distill-mxfp4_moe
96%
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
96%
think:off strata-iq3_xxs
96%
think:on strata-iq3_xxs
96%
think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
100%
think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
100%

Tool-calling (16 tasks)

think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl
33%
think:on gemma-4-26b-a4b-it-ud-q5_k_xl
75%
think:on strata-iq3_xxs
83%
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs
83%
think:off strata-swift-iq3_xxs
84%
think:off strata-iq3_xxs
86%
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
88%
think:off gemma-4-26b-a4b-it-ud-q5_k_xl
92%
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
92%
think:off qwen3_8-27b-heretic-q4_k_m
92%
think:off qwen3_8-27b-ud-q3_k_xl
92%
think:off ornith-1_5-35b-a3b-apex-mtp-quality
92%
think:off retired qwen3_8-27b-ud-q5_k_m
92%
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
92%
think:off qwen3-vl-8b-instruct-1m-q6_k
92%
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
92%
think:off retired qwen3_8-27b-ud-q4_k_m
92%
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
92%
think:off retired qwen3_8-27b-ud-q4_k_xl
92%
think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl
92%
think:off qwen36-35b-distill-mxfp4_moe
94%
think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
100%
think:on retired qwen3_8-27b-ud-q5_k_m
100%
think:on qwen3_8-27b-ud-q3_k_xl
100%
think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
100%
think:on ornith-1_5-35b-a3b-apex-mtp-quality
100%
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
100%
think:on qwen3-vl-8b-instruct-1m-q6_k
100%
think:on qwen36-35b-distill-mxfp4_moe
100%
think:on qwen3_8-27b-heretic-q4_k_m
100%

Speed (generation ms · lower is better)

think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
1707 ms
think:on qwen3_8-27b-heretic-q4_k_m
1515 ms
think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl
1408 ms
think:on gemma-4-26b-a4b-it-ud-q5_k_xl
897 ms
think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
681 ms
think:on retired qwen3_8-27b-ud-q5_k_m
665 ms
think:on qwen3_8-27b-ud-q3_k_xl
629 ms
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
613 ms
think:on strata-iq3_xxs
610 ms
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
403 ms
think:on ornith-1_5-35b-a3b-apex-mtp-quality
388 ms
think:on qwen36-35b-distill-mxfp4_moe
186 ms
think:off qwen3_8-27b-heretic-q4_k_m
67 ms
think:off strata-iq3_xxs
50 ms
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
46 ms
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
45 ms
think:off strata-swift-iq3_xxs
42 ms
think:off retired qwen3_8-27b-ud-q5_k_m
41 ms
think:off retired qwen3_8-27b-ud-q4_k_m
38 ms
think:on qwen3-vl-8b-instruct-1m-q6_k
37 ms
think:off qwen3_8-27b-ud-q3_k_xl
36 ms
think:off qwen3-vl-8b-instruct-1m-q6_k
36 ms
think:off retired qwen3_8-27b-ud-q4_k_xl
35 ms
think:off gemma-4-26b-a4b-it-ud-q5_k_xl
32 ms
think:off ornith-1_5-35b-a3b-apex-mtp-quality
29 ms
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
29 ms
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
28 ms
think:off qwen36-35b-distill-mxfp4_moe
18 ms
think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl
18 ms
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs
13 ms

Tokens per second

gen (output)

think:on qwen3_8-27b-heretic-q4_k_m
43 t/s
think:off qwen3_8-27b-heretic-q4_k_m
59 t/s
think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
63 t/s
think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
65 t/s
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
87 t/s
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
89 t/s
think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl
90 t/s
think:on retired qwen3_8-27b-ud-q5_k_m
95 t/s
think:off retired qwen3_8-27b-ud-q5_k_m
96 t/s
think:off strata-iq3_xxs
99 t/s
think:on strata-iq3_xxs
101 t/s
think:off retired qwen3_8-27b-ud-q4_k_m
105 t/s
think:on qwen3-vl-8b-instruct-1m-q6_k
106 t/s
think:on qwen3_8-27b-ud-q3_k_xl
110 t/s
think:off qwen3-vl-8b-instruct-1m-q6_k
110 t/s
think:off qwen3_8-27b-ud-q3_k_xl
110 t/s
think:off retired qwen3_8-27b-ud-q4_k_xl
112 t/s
think:off strata-swift-iq3_xxs
117 t/s
think:off gemma-4-26b-a4b-it-ud-q5_k_xl
122 t/s
think:on gemma-4-26b-a4b-it-ud-q5_k_xl
125 t/s
think:off ornith-1_5-35b-a3b-apex-mtp-quality
134 t/s
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
136 t/s
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
140 t/s
think:on ornith-1_5-35b-a3b-apex-mtp-quality
157 t/s
think:on qwen36-35b-distill-mxfp4_moe
177 t/s
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
193 t/s
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
197 t/s
think:off qwen36-35b-distill-mxfp4_moe
213 t/s
think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl
218 t/s
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs
307 t/s

prompt (input)

think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
69 t/s
think:on retired qwen3_8-27b-ud-q5_k_m
76 t/s
think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
77 t/s
think:on qwen3_8-27b-heretic-q4_k_m
77 t/s
think:on qwen3_8-27b-ud-q3_k_xl
79 t/s
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp
90 t/s
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s
94 t/s
think:off retired qwen3_8-27b-ud-q5_k_m
95 t/s
think:off qwen3_8-27b-heretic-q4_k_m
98 t/s
think:off qwen3_8-27b-ud-q3_k_xl
117 t/s
think:off retired qwen3_8-27b-ud-q4_k_xl
118 t/s
think:on strata-iq3_xxs
121 t/s
think:off retired qwen3_8-27b-ud-q4_k_m
122 t/s
think:off strata-swift-iq3_xxs
134 t/s
think:off strata-iq3_xxs
140 t/s
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs
167 t/s
think:on ornith-1_5-35b-a3b-apex-mtp-quality
173 t/s
think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl
173 t/s
think:off qwen36-35b-distill-mxfp4_moe
184 t/s
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
189 t/s
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
199 t/s
think:on qwen36-35b-distill-mxfp4_moe
200 t/s
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs
208 t/s
think:off ornith-1_5-35b-a3b-apex-mtp-quality
212 t/s
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s
214 t/s
think:on gemma-4-26b-a4b-it-ud-q5_k_xl
248 t/s
think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl
274 t/s
think:off gemma-4-26b-a4b-it-ud-q5_k_xl
278 t/s
think:on qwen3-vl-8b-instruct-1m-q6_k
568 t/s
think:off qwen3-vl-8b-instruct-1m-q6_k
576 t/s

By language (code / agent)

Modelcode:pythoncode:bashcode:javacode:typescriptagent:pythonagent:bashagent:javaagent:typescript
think:off gemma-4-26b-a4b-it-ud-q5_k_xl1.000.880.861.00————
think:on retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp1.000.870.931.001.001.001.001.00
think:on retired qwen3_8-27b-ud-q5_k_m1.000.911.000.930.831.001.001.00
think:on qwen3_8-27b-ud-q3_k_xl1.000.850.931.001.000.831.001.00
think:on qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s1.000.711.000.931.001.001.001.00
think:off qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s0.860.850.671.00————
think:off qwen36-35b-distill-mxfp4_moe0.900.690.711.00————
think:on ornith-1_5-35b-a3b-apex-mtp-quality0.930.711.001.001.000.830.831.00
think:on retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s1.000.800.930.870.830.831.001.00
think:off strata-iq3_xxs0.950.850.861.00————
think:off qwen3_8-27b-heretic-q4_k_m0.900.850.860.95————
think:on qwen3-vl-8b-instruct-1m-q6_k0.800.530.730.870.830.831.000.83
think:on retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs1.000.731.001.001.000.831.001.00
think:off qwen3_8-27b-ud-q3_k_xl1.000.690.810.81————
think:on qwen36-35b-distill-mxfp4_moe0.930.300.800.931.000.831.001.00
think:off ornith-1_5-35b-a3b-apex-mtp-quality0.810.680.711.00————
think:on qwen3_8-27b-heretic-q4_k_m1.000.710.930.930.830.830.830.83
think:off retired qwen3_8-27b-ud-q5_k_m1.000.740.670.90————
think:off retired qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp0.950.710.670.90————
think:off qwen3-vl-8b-instruct-1m-q6_k0.760.830.330.76————
think:off retired huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s0.810.680.570.95————
think:off retired qwen3_8-27b-ud-q4_k_m1.000.610.620.86————
think:on strata-iq3_xxs1.000.850.901.00————
think:off retired ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs0.760.720.710.95————
think:off retired qwen3_8-27b-ud-q4_k_xl0.900.650.671.00————
think:off retired nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs0.900.280.570.57————
think:off strata-swift-iq3_xxs1.000.680.860.95————
think:on gemma-4-26b-a4b-it-ud-q5_k_xl1.000.530.931.001.000.830.830.83
think:off retired muse-glimmer-30b-kquant-dynamic-q4_k_xl0.950.740.861.00————
think:off retired qwen3-coder-30b-a3b-instruct-ud-q4_k_xl—0.16——————

Conditions per model

gemma-4-26b-a4b-it-ud-q5_k_xl llama 0.6.0-dev+4413a19 2026-10-13
code by level: easy 100% · expert 72% · hard 85%
think: off
excluded code: 0
baseline ? (not recorded)
-c 32768
qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp llama 0.4.0-dev retired 2026-10-10
code by level: easy 97% · expert 100% · hard 88%
think: on
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 65536
qwen3_8-27b-ud-q5_k_m llama unknown retired 2026-10-10
code by level: easy 100% · expert 100% · hard 88%
think: on
excluded code: 4
baseline ? (not recorded)
--spec-type draft-mtp -c 160000 --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf
qwen3_8-27b-ud-q3_k_xl llama unknown 2026-10-10
code by level: easy 94% · expert 91% · hard 100%
think: on
excluded code: 2
baseline ? (not recorded)
--spec-type draft-mtp --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf -c 65536
qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s llama unknown 2026-10-10
code by level: easy 100% · expert 83% · hard 81%
think: on
excluded code: 1
baseline ? (not recorded)
--spec-type draft-mtp -c 200000 --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf
qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s llama 0.6.0-dev+4413a19 2026-10-13
code by level: easy 96% · expert 61% · hard 79%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 79872 --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf --spec-draft-n-max 4 --split-mode tensor
qwen36-35b-distill-mxfp4_moe llama 0.6.0-dev+4413a19 2026-10-14
code by level: easy 79% · expert 58% · hard 81%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 211968
ornith-1_5-35b-a3b-apex-mtp-quality llama unknown 2026-10-10
code by level: easy 90% · expert 100% · hard 88%
think: on
excluded code: 1
baseline ? (not recorded)
--spec-type draft-mtp -c 160000
huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s llama unknown retired 2026-10-10
code by level: easy 91% · expert 83% · hard 94%
think: on
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 160000
strata-iq3_xxs strata 0.1.43.1 ★ optimal 2026-10-13
code by level: easy 92% · expert 72% · hard 94%
think: off
excluded code: 0
baseline ? (not recorded)
qwen3.8-flash-next-iq3_xxs --pack ~/jordi/ia/llama/Strata-data/packs/iq3_xxs --native ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile ~/jordi/ia/llama/Strata/data/expert-profile.bin --expert-cache auto --vram-reserve-mib 1000 --prefill auto --spec 4 --mtp ~/jordi/ia/llama/Strata-data/mtp/rt --max-context 131072 --kv int8 --control-vector-scaled ~/jordi/ia/llama/Strata/data/experimental-speed-projection/Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer --spec-min-p 0.5 --pool-workers 4
qwen3_8-27b-heretic-q4_k_m llama 0.6.0-dev+4413a19 2026-10-13
code by level: easy 92% · expert 67% · hard 92%
think: off
excluded code: 0
baseline ? (not recorded)
--mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf -c 86016 --split-mode tensor
qwen3-vl-8b-instruct-1m-q6_k llama unknown 2026-10-10
code by level: easy 97% · expert 25% · hard 62%
think: on
excluded code: 0
baseline ? (not recorded)
-c 65536
ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs llama unknown retired 2026-10-10
code by level: easy 88% · expert 100% · hard 100%
think: on
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 160000
qwen3_8-27b-ud-q3_k_xl llama 0.6.0-dev+4413a19 2026-10-13
code by level: easy 75% · expert 58% · hard 90%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf -c 100352 --spec-draft-n-max 4
qwen36-35b-distill-mxfp4_moe llama unknown 2026-10-10
code by level: easy 93% · expert 33% · hard 88%
think: on
excluded code: 5
baseline ? (not recorded)
--spec-type draft-mtp -c 120000
ornith-1_5-35b-a3b-apex-mtp-quality llama 0.6.0-dev+4413a19 2026-10-13
code by level: easy 75% · expert 58% · hard 81%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 158720
qwen3_8-27b-heretic-q4_k_m llama unknown 2026-10-10
code by level: easy 91% · expert 91% · hard 88%
think: on
excluded code: 1
baseline ? (not recorded)
--mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf -c 65536
qwen3_8-27b-ud-q5_k_m llama 0.4.0-dev retired 2026-10-13
code by level: easy 79% · expert 67% · hard 85%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 68608 --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf --spec-draft-n-max 4 --split-mode tensor
qwen3_8-27b-coder390-efficientthink-q4lynnstyle-mtp llama 0.4.0-dev retired 2026-10-12
code by level: easy 79% · expert 61% · hard 79%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 69632 --spec-draft-n-max 4 --split-mode tensor
qwen3-vl-8b-instruct-1m-q6_k llama 0.6.0-dev+4413a19 2026-10-13
code by level: easy 100% · expert 28% · hard 62%
think: off
excluded code: 0
baseline ? (not recorded)
-c 237568
huihui-ornith-1_5-35b-a3b-abliterated_i1-iq3_s llama 0.6.0-dev+4413a19 retired 2026-10-13
code by level: easy 75% · expert 61% · hard 73%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 262144
qwen3_8-27b-ud-q4_k_m llama 0.4.0-dev retired 2026-10-12
code by level: easy 67% · expert 58% · hard 79%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf -c 84992
strata-iq3_xxs strata 0.1.43.1 2026-10-13
code by level: easy 88% · expert 86% · hard 96%
think: on
excluded code: 0
baseline ? (not recorded)
qwen3.8-flash-next-iq3_xxs --pack ~/jordi/ia/llama/Strata-data/packs/iq3_xxs --native ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf ~/jordi/ia/llama/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile ~/jordi/ia/llama/Strata/data/expert-profile.bin --expert-cache auto --vram-reserve-mib 1000 --prefill auto --spec 4 --mtp ~/jordi/ia/llama/Strata-data/mtp/rt --max-context 131072 --kv int8 --control-vector-scaled ~/jordi/ia/llama/Strata/data/experimental-speed-projection/Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer --spec-min-p 0.5 --pool-workers 4
ornith-1_5-35b-a3b-abliterated-mtpv2-ud-iq4_xs llama 0.6.0-dev+4413a19 retired 2026-10-13
code by level: easy 75% · expert 53% · hard 92%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp -c 235520
qwen3_8-27b-ud-q4_k_xl llama 0.4.0-dev retired 2026-10-12
code by level: easy 71% · expert 58% · hard 85%
think: off
excluded code: 0
baseline ? (not recorded)
--spec-type draft-mtp --mmproj ~/jordi/ia/llama/models/mmproj-F16.gguf -c 79872
nvidia-nemotron-3_5-lightning-30b-a3b-iq4_xs llama 0.4.0-dev retired 2026-10-12
code by level: easy 26% · expert 33% · hard 81%
think: off
excluded code: 6
baseline ? (not recorded)
--spec-type draft-mtp -c 364544 --reasoning-format none
strata-swift-iq3_xxs strata 0.1.44.1 2026-10-14
code by level: easy 75% · expert 75% · hard 83%
think: off
excluded code: 0
baseline ? (not recorded)
swift-1.5-iq3_xxs --pack ~/jordi/ia/llama/Strata-data/packs/swift-iq3_xxs --native ~/jordi/ia/llama/Strata-data/models/HUIHUI-V2-IQ3_XXS/Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf ~/jordi/ia/llama/Strata-data/models/HUIHUI-V2-IQ3_XXS/Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --expert-profile ~/jordi/ia/llama/Strata/data/expert-profile.bin --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp ~/jordi/ia/llama/Strata-data/mtp/rt --max-context 131072 --kv int8
gemma-4-26b-a4b-it-ud-q5_k_xl llama unknown 2026-10-10
code by level: easy 84% · expert 83% · hard 94%
think: on
excluded code: 0
baseline ? (not recorded)
-c 100000
muse-glimmer-30b-kquant-dynamic-q4_k_xl llama 0.4.0-dev retired 2026-10-12
code by level: easy 75% · expert 81% · hard 92%
think: off
excluded code: 0
baseline ? (not recorded)
-c 131072 -md ~/jordi/ia/llama/models/dflash-Muse-Glimmer-30B-Q4_K_M.gguf --mmproj ~/jordi/ia/llama/models/mmproj-Muse-Glimmer-30B-Q4_K_M.gguf
qwen3-coder-30b-a3b-instruct-ud-q4_k_xl llama 0.4.0-dev retired 2026-10-12
code by level: easy 16% · expert 0% · hard 50%
think: off
excluded code: 29
baseline ? (not recorded)
-c 215040

Recommendations (by use)

gemma-4-26b-a4b-it-ud-q5_k_xl think:off

0.84
global · llama 0.6.0-dev+4413a19 · 2026-10-13
91% · 93% · 92%
code · agent · tools (excluded code: 0)
32 ms · 122 t/s
latency · throughput
  • Best overall — daily work in opencode (balance)

qwen3_8-27b-ud-q3_k_xl think:on

0.83
global · llama unknown · 2026-10-10
95% · 96% · 100%
code · agent · tools (excluded code: 2)
629 ms · 110 t/s
latency · throughput
  • Best code — hard/expert code tasks
  • Best tools — tool-calling and orchestration
  • Best thinking — hard problems where thinking pays off the latency

qwen3_8-27b-turbofcfusion-735-882-here-uncen-neo-coder-max-mtp-q4_k_s think:on

0.83
global · llama unknown · 2026-10-10
92% · 100% · 100%
code · agent · tools (excluded code: 1)
1707 ms · 63 t/s
latency · throughput
  • Best agent — autonomous multi-step work

qwen36-35b-distill-mxfp4_moe think:off

0.81
global · llama 0.6.0-dev+4413a19 · 2026-10-14
76% · 77% · 94%
code · agent · tools (excluded code: 0)
18 ms · 213 t/s
latency · throughput
  • Fastest — interactivity (lowest latency)
  • Highest throughput — long generations