Freshness / lifecycle. The current release is hipfire 0.4.3.1; optimization continues on the beta branch, so builds newer than the release can differ from the numbers here. Every figure below is fixture-scoped, and a date is shown only when its source carries a measurement date. These source-published snapshots are not lasting claims about every build. /docs/benchmarks is the build-time live ledger.
Rows are heterogeneous fixtures. Compare cells only when model, quant/mode, prompt, method, and backend match. Engine sources and methodology docs point at warpfront/hipfire (the current release, v0.4.4).
hipfire 0.4.4. The current release is 0.4.4 (2026-10-27). Its numbers (Qwen3.8-27B DFlash and MTP on the R9700, 7900 XTX and Strix Halo, Flash-Next on the R9700 and the Strix Halo NPU, A3B mq4rs, dense TP2/TP4) live on the 0.4.4 benchmarks page with their fixtures and sources. Earlier releases keep their pages: 0.4.3 (older release), 0.4.2 and 0.4.1. The featured snapshots below are from the 0.4.3 and 0.4.3.1 releases, an older release. The historical snapshots at the end of the featured section predate 0.4.1.
Last build: 2026-10-29T04:56:45.779Z ·
live source: GET /api/leaderboard?engineName=hipfire ·
engine: github.com/warpfront/hipfire current release v0.4.4; optimization continues on branch beta
featured snapshots · 0.4.3 and 0.4.3.1
Compact tables copied from the 0.4.3 and 0.4.3.1 release notes. Each block names its date, card, fixture and method, and links the full page, so the figures stay scoped to that measurement, not a lasting claim about every hipfire build.
Qwen3.6 35B-A3B mq4rs · Radeon AI PRO R9700 (gfx1201) · prefill, 2026-10-24
| prefill tok/s | 0.4.2.1 mq4r | 0.4.3 mq4r | 0.4.3 mq4rs | mq4rs vs 0.4.2.1 mq4r |
|---|---|---|---|---|
| pp512 | 3,292 | 3,332 | 11,589 | 3.5× |
| pp1024 | 3,248 | 3,607 | 15,718 | 4.8× |
| pp4096 | 2,886 | 3,737 | 19,828 | 6.9× |
| pp8192 | 2,488 | 3,615 | 20,035 | 8.1× |
| pp16384 | 1,957 | 3,561 | 18,545 | 9.5× |
Radeon AI PRO R9700, hipfire bench --matrix --pp 512,1024,4096,8192,16384 --tg 1 --ctx 1 --runs 3 --warmups 1 --spec off,
median of 9 samples (3 fresh processes), arms interleaved in one timing hold, 2026-10-24. “0.4.2” is v0.4.2.1.
Each version ran its default config and the defaults differ (KV q8 and 512-row chunks on 0.4.2.1; fp8 KV and
8192-row chunks on 0.4.3), so the mq4r-to-mq4r column is an upgrade comparison, not a single-variable A/B.
Fast mq4rs prefill is gfx1201 only. Source: v0.4.3 release notes; details on the
0.4.3 page.
Qwen3.6 35B-A3B mq4rs · Radeon AI PRO R9700 (gfx1201) · decode, 0.4.3.1
| decode tok/s | HIP route (0.4.3) | PM4 replay (0.4.3.1) | change |
|---|---|---|---|
| tg128, ctx 128 | 157 | 232 | +48 % |
| tg128, ctx 8K | 146.6 | 217.0 | +48 % |
| 1024-token reasoning generation | 154 | 225.0 | +46 % |
Radeon AI PRO R9700, hipfire bench --spec off --matrix --pp 128 --ctx 128,8192 --tg 128 --runs 3 --warmups 2, default
config (HIP arm set with replay.backend = "hip"), 3 fresh processes per arm; greedy ids identical
between routes. gfx1201 only; prefill unaffected. Source: v0.4.3.1 release notes;
more detail on the 0.4.3 page.
Qwen3.8-Flash (Flash-Next) · R9700 prefill and Strix Halo 16K
| change | fixture | result |
|---|---|---|
| MQ6 INT8 prefill, now default on gfx1201 | One R9700 · Flash-Next · pp16384 at the 4096-row chunk · 3 fresh processes per arm | 4,720.5 → 4,899.3 tok/s (+3.8 %) against the previous F16 route |
| Compact forward scratch, now default on gfx1201 (4096-row chunk) | One Radeon AI PRO R9700 · split base qwen3.8-flash-next-base.mq4-xts · hipfire bench --matrix pp16384, tg128 and --spec auto · 3 fresh processes per arm · full vs compact scratch, both at the 4096-row chunk | Forward scratch 4,388 → 1,270 MiB; resident experts 6,948 → 8,234. pp16384 5,024 → 5,405 tok/s (+7.6 %); tg128 33.56 → 34.84 (+3.8 %); --spec auto prose 45.8 → 49.0 (+7.0 %), code 68.2 → 73.9 (+8.4 %) |
| Default prefill chunk is now 16,384 rows on gfx1201: faster prefill, slower decode | One Radeon AI PRO R9700 · gptq3 monolith · pp16384, AR tg128 (synthetic random token ids, understates real-text decode) and --spec auto (fiction_lighthouse prose, humaneval_0_has_close_elements code) · 3 fresh processes per arm · 16,384-row vs 4096-row chunk, both with compact scratch | pp16384 5,405 → 6,431 tok/s (+19.0 %). Decode costs: AR tg128 34.88 → 33.11 (−5.1 %); --spec auto prose 49.1 → 44.3 (−9.8 %), code 73.7 → 66.1 (−10.3 %); resident experts 8,234 → 6,578. HIPFIRE_PREFILL_CHUNK_ROWS=4096 restores the old decode speed |
| Four exact-kernel prefill waves W1–W4 (default-on) | One Radeon AI PRO R9700 · pp16384 at the 16,384-row chunk with compact scratch · 3 fresh processes per arm · each wave an interleaved same-session A/B | W1 5,506.8 → 5,926.1 (+7.6 %); W2 5,903.9 → 6,028.8 (+2.1 %); W3 6,037.2 → 6,385.5 (+5.8 %); W4 6,380.1 → 6,449.8 (+1.1 %). Measured in separate sessions, not chained into one total |
| Strix Halo 16K prefill, NPU trunk on | Strix Halo + XDNA2 · 16,355-token prompt · client TTFT · 3 fresh processes × 3 requests · 2026-10-23 | 2,365.6 tok/s all-9 median; four steady requests 2,403–2,407. v0.4.2 shipped tip, an earlier measurement (2026-10-14): 2,386 tok/s |
Rows are separate change-level campaigns on different cards and are not chained. Source:
v0.4.3 release notes, Headline numbers. The 16,384-row chunk makes decode 5–10 %
slower; HIPFIRE_PREFILL_CHUNK_ROWS=4096 restores it. The expert hot set, MTP figures published
with the release notes and the Strix Halo contributor figures are on the
0.4.3 page.
historical snapshots (before 0.4.1)
The three snapshots below come from the engine's README and perf-checkpoint files and predate the 0.4.1
release matrix. They are kept as dated evidence of those fixtures, not as current claims, and were not
re-measured for 0.4.3. In particular the Qwen3.6 35B-A3B row is the mq4r file, not
qwen3.6:35b-a3b-mq4rs, so it is not a baseline for the 0.4.3.1 decode numbers above.
Qwen3.6 35B-A3B MQ4R · single-GPU AR · historical README snapshot
Ordinary autoregressive decode, single GPU; no MTP, DFlash, reduced-output bench, or manual clock pinning. TG128 = three-run medians. Multi-turn columns from clean eight-turn serving runs.
Historical: an mq4r AR snapshot from the engine README, with no measurement date recorded here, not re-measured
for 0.4.3. The R9700 row (203.9 tok/s) is a different file from the 0.4.3.1 mq4rs decode figure above; the
7900 XTX and Strix Halo rows have no 0.4.3 re-measurement.
| GPU | arch | TG128 AR | 8-turn avg | final turn |
|---|---|---|---|---|
| Radeon RX 7900 XTX | gfx1100 | 253.3 tok/s | 191.0 tok/s | 160.3 @ 18.2K |
| Radeon 8060S / Strix Halo | gfx1151 | 115.1 tok/s | 92.2 tok/s | 82.5 @ 21.3K |
| Radeon AI PRO R9700 | gfx1201 | 203.9 tok/s | 169.5 tok/s | 146.7 @ 22.2K |
Source: README · Qwen3.6 35B-A3B MQ4R
Qwen3.8-27B MQ4V2 · 2026-08-20 · hiptrx · gfx1201 · Radeon AI PRO R9700 · historical
Product-ladder checkpoint on the hiptrx gfx1201 / Radeon AI PRO R9700 fixture. AR decode, prefill, DFlash decode, mean accepted draft length (τ), and bits-per-weight.
Historical: a 2026-08-20 perf checkpoint on the old XT / Base / Pro product ladder. It is dated evidence for that build and fixture, not a current claim, and it predates the 0.4.1 release numbers.
| tier | AR decode | prefill | DFlash | τ | bpw |
|---|---|---|---|---|---|
| XT | 35.3 tok/s | 490.7 | 251.6 tok/s | 11.70 | 4.456 |
| Base | 33.2 tok/s | 479.0 | 263.3 tok/s | 13.11 | 4.659 |
| Pro | 31.8 tok/s | 473.8 | 258.2 tok/s | 13.11 | 4.897 |
Dated 2026-08-20. Source:
2026-08-20 Qwen3.8 MQ-V2 product ladder
(docs/perf-checkpoints/2026-08-20-qwen38-mq-v2-product-ladder.md)
DeepSeek V4 Flash MQ2R · 4× Radeon AI PRO R9700 (gfx1201) · historical
n=3 fresh-process medians; greedy; speculative decode off; KV f32; 2052-token prompt.
Historical README snapshot, not re-measured for 0.4.3.
| topology | decode | prefill |
|---|---|---|
| TP3/EP | 53.1 tok/s | 481 tok/s |
| TP4/EP | 54.3 tok/s | 389 tok/s |
Source: README · RDNA4 gfx1201 R9700
live localmaxxing ledger
Architecture tables below are approved community runs pulled at this site build. They are a separate surface from the featured snapshots above — different harness, different fixtures. Do not treat a live cell and a dated snapshot as the same measurement.
gfx1100 — RDNA 3
| hardware | model | q | mode | decode tok/s | prefill | AR base | × AR | τ | run |
|---|---|---|---|---|---|---|---|---|---|
| RX 7900 XTX | LFM2.5-350M | MQ4 | AR | 1470.8 | — | — | — | — | cmsnp1w730… |
| RX 7900 XTX | LFM2.5-350M | MQ4 | AR | 885.3 | — | — | — | — | cmsnp2nkg0… |
| RX 7900 XTX | LFM2.5-350M | MQ4 | AR | 858.4 | — | — | — | — | cmsnp2opb0… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | DFlash | 576.9 | 755 | 122.9 | 4.70× | 13.18 | cmpdzs3fq0… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | DFlash | 575.3 | — | — | — | — | cmofzk74w0… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | DFlash | 575.2 | — | 123.1 | 4.67× | 13.18 | cmp8fqdz20… |
| RX 7900 XTX | Qwen3.5-0.8B | MQ4 | AR | 550.2 | — | — | — | — | cmsnp33j90… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4R | MTP | 494.6 | 1468 | — | — | — | cmsnp3cmd0… |
| RX 7900 XTX | LFM2.5-350M | MQ4 | AR | 458.7 | — | — | — | — | cmsnp3h5g0… |
| RX 7900 XTX | LFM2.5-350M | MQ4 | AR | 398.4 | — | — | — | — | cmsnp3mx40… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | DFlash | 337.0 | — | — | — | — | cmofdp1ve0… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | DFlash | 322.6 | — | — | — | — | cmofkl1gm0… |
| RX 7900 XTX | Qwen3.8-27B | MQ4V2 | DFlash | 318.2 | 338 | — | — | — | cmu26wlwc0… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4R | AR | 272.6 | — | — | — | — | cmsnp41sa0… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4R | AR | 270.8 | — | — | — | — | cmsnp42xp0… |
| RX 7900 XTX | Qwen3.5-27B | MQ4 | DFlash | 254.4 | — | 44.6 | 5.70× | 13.18 | cmp8fozao0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | DFlash | 252.1 | 367 | — | — | — | cms8csy5d0… |
| RX 7900 XTX | Qwen3.5-27B | MQ4 | DFlash | 250.3 | — | — | — | — | cmofyatrj0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4-AWQ | DFlash | 232.1 | 335 | — | — | — | cmppe779e0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | DFlash | 216.7 | — | 44.8 | 4.84× | 10.93 | cmp8fw36n0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | DFlash | 216.6 | — | 44.6 | 4.85× | 10.93 | cmp8fnkw00… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4R | AR | 214.7 | 913 | — | — | — | cmsnp49si0… |
| RX 7900 XTX | Qwen3.5-27B | MQ4 | DFlash | 201.1 | — | — | — | — | cmofkrpp00… |
| RX 7900 XTX | Qwen3.5-27B | MQ4 | DFlash | 182.0 | — | — | — | — | cmofe2efg0… |
| RX 7900 XTX | Qwen3.5-35B-A3B | MQ4 | DFlash | 154.6 | — | — | — | — | cmofkye0g0… |
| RX 7900 XTX | Qwen3.8-27B | MQ4 | DFlash | 154.2 | 51 | — | — | — | cmt1ff27w0… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4R | AR | 150.0 | — | — | — | — | cmsnp4ebk0… |
| RX 7900 XTX | Qwen3.5-35B-A3B | MQ4 | DFlash | 140.6 | — | — | — | — | cmofefr0j0… |
| RX 7900 XTX | Qwen3.5-35B-A3B | MQ4 | AR | 135.9 | — | — | — | — | cmofe92tt0… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4 | AR | 135.3 | — | — | — | — | cmofezrun0… |
| RX 7900 XTX | Ornstein3.6-35B-A3B | MQ4 | AR | 134.0 | — | — | — | — | cmofgd9iz0… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | AR | 123.1 | — | — | — | — | cmp8ft6rt0… |
| RX 7900 XTX | Qwen3.5-9B | MQ4 | AR | 122.3 | — | — | — | — | cmofdidib0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | DFlash | 118.2 | — | — | — | — | cmofet3j80… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | DFlash | 118.1 | — | — | — | — | cmofl528t0… |
| RX 7900 XTX | Qwen3.6-35B-A3B | MQ4 | DFlash | 68.6 | — | — | — | — | cmoff6g560… |
| RX 7900 XTX | Qwen3.8-27B-GGUF | MQ4 | AR | 46.2 | 1012 | — | — | — | cmsv00obw0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | AR | 44.8 | — | — | — | — | cmp8frsd70… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | AR | 44.6 | — | — | — | — | cmp8ful710… |
| RX 7900 XTX | Qwen3.5-27B | MQ4 | AR | 44.6 | — | — | — | — | cmp8fl9rq0… |
| RX 7900 XTX | Qwen3.5-27B | MQ4 | AR | 43.7 | — | — | — | — | cmofdvq7c0… |
| RX 7900 XTX | Qwen3.6-27B | MQ4 | AR | 43.6 | — | — | — | — | cmofemf840… |
Engine: hipfire@0.3.0+b0bcc3f91506 ·
Submitter: @schuttdev
gfx1201 — RDNA 4
| hardware | model | q | mode | decode tok/s | prefill | AR base | × AR | τ | run |
|---|---|---|---|---|---|---|---|---|---|
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 6001.4 | — | — | — | — | cmsnp1rko0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 5993.3 | — | — | — | — | cmsnp232a0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 4816.1 | — | — | — | — | cmsnp1spu0… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 4347.3 | — | — | — | — | cmsnp1tv30… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 4207.7 | — | — | — | — | cmsnp24750… |
| RX 9070 XT | LFM2.5-350M | MQ4 | AR | 3669.8 | — | — | — | — | cmsnp1v0f0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 3283.3 | — | — | — | — | cmsnp25c40… |
| RX 9070 XT | LFM2.5-350M | MQ4 | AR | 1876.9 | — | — | — | — | cmsnp26gr0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1704.6 | — | — | — | — | cmsnp27le0… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 1597.7 | — | — | — | — | cmsnp28q50… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 1585.6 | — | — | — | — | cmsnp29vb0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1523.1 | — | — | — | — | cmsnp2b0g0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1457.0 | — | — | — | — | cmsnp2c5d0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1376.7 | — | — | — | — | cmsnp2d9x0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 1291.9 | — | — | — | — | cmsnp2eef0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 1278.9 | — | — | — | — | cmsnp2fk00… |
| RX 9070 XT | LFM2.5-350M | MQ4 | AR | 1258.5 | — | — | — | — | cmsnp2gp30… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1246.8 | — | — | — | — | cmsnp2hup0… |
| RX 9070 XT | LFM2.5-350M | MQ4 | AR | 1237.7 | — | — | — | — | cmsnp2j030… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 1115.7 | — | — | — | — | cmsnsxbrr0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1044.8 | — | — | — | — | cmsnp2k4r0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 1017.0 | — | — | — | — | cmsnp2l9o0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 969.4 | — | — | — | — | cmsnsxd8w0… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 597.8 | — | — | — | — | cmsnp2s4l0… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 597.2 | — | — | — | — | cmsnp2t9n0… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 596.6 | — | — | — | — | cmsnp2ue80… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 594.1 | — | — | — | — | cmsnp2viy0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 574.3 | — | — | — | — | cmsnp2wnr0… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 567.5 | — | — | — | — | cmsnp2xsx0… |
| Radeon AI Pro R9700 | Qwen3.5-0.8B | MQ4 | AR | 563.5 | — | — | — | — | cmsnp20sc0… |
| Radeon AI Pro R9700 | Qwen3.5-0.8B | MQ4 | AR | 562.4 | — | — | — | — | cmsnp2yyv0… |
| Radeon AI Pro R9700 | Qwen3.5-0.8B | MQ4 | AR | 562.0 | — | — | — | — | cmsnp30410… |
| Radeon AI Pro R9700 | Qwen3.5-0.8B | MQ4 | AR | 554.5 | — | — | — | — | cmsnp32eb0… |
| RX 9070 XT | Qwen3.5-0.8B | MQ4 | AR | 545.3 | — | — | — | — | cmsnp34or0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 542.5 | — | — | — | — | cmsnp35tm0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 542.1 | — | — | — | — | cmsnp36y20… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 539.6 | — | — | — | — | cmsnp382m0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 538.3 | — | — | — | — | cmsnp397g0… |
| RX 9070 XT | LFM2.5-350M | MQ4 | AR | 520.8 | — | — | — | — | cmsnp3ac90… |
| Radeon AI Pro R9700 | LFM2.5-230M | MQ4 | AR | 520.0 | — | — | — | — | cmsnp3bh60… |
| RX 9070 XT | LFM2.5-230M | MQ4 | AR | 481.8 | — | — | — | — | cmsnp3dqp0… |
| Radeon AI Pro R9700 | LFM2.5-350M | MQ4 | AR | 466.0 | — | — | — | — | cmsnp3g0m0… |
| RX 9070 XT | LFM2.5-350M | MQ4 | AR | 437.7 | — | — | — | — | cmsnp3jic0… |
| Radeon AI Pro R9700 | Qwen3.5-9B | MQ4 | DFlash | 371.8 | 1160 | 99.4 | 3.74× | 13.18 | cmpdzu8ig0… |
| Radeon AI Pro R9700 | Qwen3.8-27B | MQ4V2 | DFlash | 303.4 | 472 | 36.7 | — | — | cmu25i4qc0… |
| Radeon AI Pro R9700 | Qwen3.5-27B | MQ4-AWQ | DFlash | 286.6 | 436 | — | — | — | cmppftgt80… |
| Radeon AI Pro R9700 | Qwen3.8-27B | INT4 | DFlash | 286.3 | 435 | — | — | — | cmt62jgz20… |
| Radeon AI Pro R9700 | Qwen3.8-27B | MQ4V2 | DFlash | 282.4 | 451 | — | — | — | cmt0pjfxq0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | MTP | 280.4 | 1469 | — | — | — | cmsnp3tqo0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | MTP | 279.4 | 1469 | — | — | — | cmsnp3uvq0… |
| Radeon AI Pro R9700 | Qwen3.6-27B | MQ4 | DFlash | 278.0 | 467 | — | — | — | cms8cr7d70… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | MTP | 277.4 | 1469 | — | — | — | cmsnp3x5g0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | MTP | 274.9 | 1469 | — | — | — | cmsnp3zfg0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4-AWQ | MTP | 259.5 | 1094 | 259.5 | — | — | cmppmb47l0… |
| Radeon AI Pro R9700 | Qwen3.6-27B | MQ4-AWQ | DFlash | 254.8 | 435 | — | — | — | cmppfddi30… |
| Radeon AI Pro R9700 | LFM2.5-8B-A1B | MQ4 | AR | 241.1 | 242 | — | — | — | cmpruarq60… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 233.6 | — | — | — | — | cmsnp46b50… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 230.9 | — | — | — | — | cmsnp47g50… |
| Radeon AI Pro R9700 | Qwen3.6-27B | MQ4 | SPEC | 213.7 | 750 | — | — | — | cmrgp8xgt0… |
| Radeon AI Pro R9700 | Qwen3.5-27B | MQ4 | DFlash | 196.2 | 438 | 35.4 | 5.54× | 13.18 | cmpdzvqpx0… |
| Radeon AI Pro R9700 | Qwen3.6-35B-A3B | MQ4R | AR | 127.3 | — | — | — | — | cmsnp4fgp0… |
| Radeon AI Pro R9700 | Qwen3.8-27B | MQ4V2 | DFlash | 111.6 | 721 | — | — | — | cmu25i4g80… |
| Radeon AI Pro R9700 | DeepSeek-V4-Flash | MQ2R | AR | 54.3 | 389 | — | — | — | cmsnp21x40… |
| Radeon AI Pro R9700 | DeepSeek-V4-Flash | MQ2R | AR | 53.1 | 481 | — | — | — | cmsnp4iw40… |
Engine: hipfire@0.3.0+b0bcc3f91506 ·
Submitter: @schuttdev
gfx1151 — RDNA 3.5
| hardware | model | q | mode | decode tok/s | prefill | AR base | × AR | τ | run |
|---|---|---|---|---|---|---|---|---|---|
| Strix Halo (Radeon 8060S Graphics) | Qwen3.5-9B | MQ4 | DFlash | 255.8 | 406 | 45.6 | 5.60× | 13.18 | cmpdzx93g0… |
| Strix Halo (Radeon 8060S Graphics) | Qwen3.5-27B | MQ4 | DFlash | 104.5 | 135 | — | — | 13.18 | cmpe035k30… |
| Strix Halo (Radeon 8060S Graphics) | Qwen3.6-27B | MQ4 | DFlash | 88.3 | 134 | 14.8 | 5.96× | 10.93 | cmpe04nu50… |
| Radeon 8060S | DeepSeek-V4-Flash | MQ2-Lloyd + Q8 KV + MTP sidecar | SPEC | 19.0 | 146 | — | — | — | cmpqeql2m0… |
| Radeon 8060S | Qwen3.5-27B | MQ4 | AR | 14.8 | — | — | — | — | cmoknenfb0… |
Engine: hipfire@0.2.0+1a378379 ·
Submitter: @schuttdev
gfx1030 — RDNA 2
| hardware | model | q | mode | decode tok/s | prefill | AR base | × AR | τ | run |
|---|---|---|---|---|---|---|---|---|---|
| RX 6900 XT | LFM2.5-350M | MQ4 | AR | 584.2 | — | — | — | — | cmsnp1znf0… |
| RX 6900 XT | LFM2.5-350M | MQ4 | AR | 557.2 | — | — | — | — | cmsnp319s0… |
| RX 6900 XT | Qwen3.5-0.8B | MQ4 | AR | 346.1 | — | — | — | — | cmsnp3qbh0… |
| RX 6900 XT | LFM2.5-350M | MQ4 | AR | 275.8 | — | — | — | — | cmsnp3yab0… |
| RX 6950 XT | Qwen3.5-9B | MQ4 | DFlash | 222.0 | 479 | 75.1 | 2.96× | 13.18 | cmpe01n920… |
Engine: hipfire@0.3.0+b0bcc3f91506 ·
Submitter: @schuttdev
gfx1010 — RDNA 1
| hardware | model | q | mode | decode tok/s | prefill | AR base | × AR | τ | run |
|---|---|---|---|---|---|---|---|---|---|
| RX 5700 XT | LFM2.5-230M | MQ4 | AR | 648.3 | — | — | — | — | cmsnp1yij0… |
| RX 5700 XT | LFM2.5-230M | MQ4 | AR | 432.3 | — | — | — | — | cmsnp3kn50… |
| RX 5700 XT | LFM2.5-230M | MQ4 | AR | 428.0 | — | — | — | — | cmsnp3ls00… |
| RX 5700 XT | LFM2.5-350M | MQ4 | AR | 394.0 | — | — | — | — | cmsnp3o210… |
| RX 5700 XT | LFM2.5-350M | MQ4 | AR | 390.8 | — | — | — | — | cmsnp3p6o0… |
| RX 5700 XT | LFM2.5-230M | MQ4 | AR | 279.3 | — | — | — | — | cmsnp3w0e0… |
| RX 5700 XT | Qwen3.5-0.8B | MQ4 | AR | 273.6 | — | — | — | — | cmsnp40li0… |
| RX 5700 XT | LFM2.5-350M | MQ4 | AR | 257.3 | — | — | — | — | cmsnp44230… |
| RX 5700 XT | LFM2.5-230M | MQ4 | AR | 226.3 | — | — | — | — | cmsnp48nj0… |
| RX 5700 XT | LFM2.5-350M | MQ4 | AR | 207.7 | — | — | — | — | cmsnp4ax10… |
| RX 5700 XT | Qwen3.5-9B | MQ4 | AR | 61.6 | 210 | — | — | — | cmpdzyrb70… |
Engine: hipfire@0.3.0+b0bcc3f91506 ·
Submitter: @schuttdev
? — ?
| hardware | model | q | mode | decode tok/s | prefill | AR base | × AR | τ | run |
|---|---|---|---|---|---|---|---|---|---|
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-350M | MQ4 | AR | 1181.6 | — | — | — | — | cmsnp1xcx0… |
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-230M | MQ4 | AR | 958.0 | — | — | — | — | cmsnp2mfn0… |
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-350M | MQ4 | AR | 777.0 | — | — | — | — | cmsnp2pua0… |
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-350M | MQ4 | AR | 766.3 | — | — | — | — | cmsnp2qzh0… |
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-230M | MQ4 | AR | 470.3 | — | — | — | — | cmsnp3ew10… |
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-350M | MQ4 | AR | 439.1 | — | — | — | — | cmsnp3ian0… |
| Ryzen AI Max (Ryzen AI Max 395+) | LFM2.5-350M | MQ4 | AR | 343.1 | — | — | — | — | cmsnp3rgh0… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.5-0.8B | MQ4 | AR | 291.4 | — | — | — | — | cmsnp3slf0… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.6-35B-A3B | MQ4R | MTP | 248.2 | 1468 | — | — | — | cmsnp456p0… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.6-35B-A3B | MQ4R | AR | 161.8 | — | — | — | — | cmsnp4c290… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.6-35B-A3B | MQ4R | AR | 161.4 | — | — | — | — | cmsnp4d790… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.6-27B | MQ4 | DFlash | 112.0 | 159 | — | — | — | cms8cugd00… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.8-27B | MQ4V2 | DFlash | 104.3 | 144 | — | — | — | cmu26wm6u0… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.8-27B | MQ4V2 | DFlash | 103.9 | 144 | — | — | — | cmt6wvmy40… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.6-35B-A3B | MQ4R | AR | 101.7 | 536 | — | — | — | cmsnp4glp0… |
| Ryzen AI Max (Ryzen AI Max 395+) | Qwen3.6-35B-A3B | MQ4R | AR | 91.4 | — | — | — | — | cmsnp4hqe0… |
| Ryzen AI Max (Ryzen AI Max 395+) | DeepSeek-V4-Flash | MQ2R | SPEC | 39.0 | — | — | — | — | cmsnu5ajq0… |
| Ryzen AI Max (Ryzen AI Max 395+) | DeepSeek-V4-Flash | MQ2R | AR | 30.8 | 202 | — | — | — | cmsnp4k0u0… |
| Strix Halo (Ryzen AI Max+ 395) | Qwen3.6-27B | MQ4 | AR | 17.0 | — | — | — | — | cmokadzkk0… |
| Ryzen AI Max (Ryzen AI Max 395+) | DeepSeek-V4-Flash | MQ2-Lloyd | MTP | 15.2 | 145 | — | — | — | cmrh7bdda0… |
Engine: hipfire@0.3.0+b0bcc3f91506 ·
Submitter: @schuttdev
methodology
Live localmaxxing cells follow a shared protocol so cross-arch comparisons inside that harness are meaningful. Featured snapshots use the method notes printed beside each table — those fixtures are not interchangeable with the live ledger.
- Prompt:
benchmarks/prompts/merge_sort_thinking_off.txt, md5253c7ac50857fe6d0e10fb0d2c5e35c0, 27 input tokens. One newline-shape variant locks in a stable τ across runs. Source tree: warpfront/hipfire. - Config:
--max 256 --temp 0.0 --no-chatml --kv-mode q8 --ctx 4096,prompt_normalize=on(default since 2026-04-26). - Runs: one
--max 16warmup per cell (discarded), then 3 timed runs at--max 256, median reported. AR baseline is a paired--ar-baselinerun from the same binary on the same hardware. - Coherence: every submitted run produced readable
merge_sortcode — no attractors, token loops, or special-token leaks. - GPU isolation:
HIP_VISIBLE_DEVICESper cell on hosts with multiple GPUs; no two cells share a GPU concurrently.
Why this matters: a single newline change in the prompt can swing τ by 17% on 27B DFlash — same model, same flags, different token sequence. The prompt md5 is part of the claim.
Engine docs and performance checkpoints live in
warpfront/hipfire (current release v0.4.4; optimization continues on
beta) — see
docs/ and
docs/perf-checkpoints/.
reproducibility
Every live row carries the exact launch command in its localmaxxing notes. To repro any cell:
$ hipfire pull <model> $ hipfire bench <model> --spec dflash --runs 5 --warmups 3 \ --max-tokens 256 --backend noslots --workload stateless --json
Prefer hipfire bench over examples/dflash_spec_demo:
the demo runs raw open-ended continuation and under-reports by ~40%
because it never hits the chatml + structured-thinking serve path
(see AGENTS.md).
For DFlash, the draft model is hosted at
z-lab/Qwen3.5-9B-DFlash
and
z-lab/Qwen3.5-27B-DFlash.
The native-head MQ4 conversion is shipped with hipfire's draft pulls
(hipfire pull qwen3.5:9b-dflash).
submit your own
Anyone with a hipfire build can submit. Install
lmx-bench for a curated harness, or POST directly to
localmaxxing.com/api/benchmarks with the schema from
the API docs.
See an existing row's engineFlags.commandSnippet for the
exact invocation to mirror. Build from
warpfront/hipfire (the release tag, or beta for in-progress work).