FastLLM Preview 2026.10.21.12 test results

Release checks

Pass 14 checks on the release. Covers: SHA256SUMS, installer and package signatures, signed file catalog, installed tree equals package, uninstall keeps data.

Installer: FastLLM-Setup-2026.10.21.12.exe
SHA-256 f96916b08caf5b06ecef51a688eac65621a925a2f566625aa86328a90182eb95

Release page on GitHub

Test hardware

AMD Radeon RX 7900 XT 20GB — Resizable BAR on
AMD Radeon RX 7900 XTX 24GB — Resizable BAR off

Automated tests

Check	Result	Cards	Account
Update from 2026.10.21.11 (canary) to 2026.10.21.12 over the stable channel with FastLLM's own updater Both PCs: the update command over the stable channel from the canary 2026.10.21.11 (same commit as 2026.10.21.12) to 2026.10.21.12. SHA256SUMS, installer size, hash, Bitworks signature and version verified; installed files match the signed catalog (install integrity ok).	Pass	XT, XTX	standard Windows user
Window setup check, driver guidance, setup list and preview read through the window's own output capture	Pass	XT, XTX	standard Windows user
Default start picks the HIP setup, no fallback, self-test passes	Pass	XT, XTX	standard Windows user
Vulkan setup starts when chosen, self-test passes	Pass	XT, XTX	standard Windows user
Nothing downloaded during starts; no new driver or application errors	Pass	XT, XTX	standard Windows user

Self-test speed

Workload: 1 discarded warm-up + 2 measured runs; 512-token prompt; 128 generated tokens; server timings (median)

Card	Setup	Generation (tok/s)	Prompt (tok/s)	First token
AMD Radeon RX 7900 XT 20GB	HIP (the setup FastLLM picks)	33.3	811.6	633 ms
AMD Radeon RX 7900 XT 20GB	Vulkan (fallback)	33.9	532.1	964 ms
AMD Radeon RX 7900 XTX 24GB	HIP (the setup FastLLM picks)	33.5	870.8	590 ms
AMD Radeon RX 7900 XTX 24GB	Vulkan (fallback)	35.8	658.4	779 ms

HIP processes the prompt much faster and starts answering sooner; generation speed is about the same as Vulkan.

Window checklist

Real FastLLM window operated by remote desktop: on the RX 7900 XTX PC with an administrator Windows account (User Account Control on; FastLLM runs without elevation), and on the RX 7900 XT PC with a standard Windows account (lines 3, 6, 8, 10 to 13 and 16 on 2026.10.15.12; lines 3, 12, 13, 16 and 17 to 19 on 2026.10.15.13).

“Ran on” lists the releases each line was run on. A line not re-run on 2026.10.21.12 keeps the earlier release it passed on.

Line	Check	Result	Notes	Ran on
1	Download the signed installer from the release page and install it	Pass	No admin prompt. Edge may hold the file as not commonly downloaded; Keep releases it. Publisher shown as Bitworks LLC.	2026.10.15.9, 2026.10.15.11
3	Open FastLLM	Pass	Title and notice correct; no integrity warning. Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.9 to .13
6	Plan names the best setup for the card	Pass	"Measured by Bitworks on this card" with a See results link. Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.9, 2026.10.15.12
7	Model license shown before any download	Pass	Exact model, revision, license and size reviewed before a 15.33 GiB download; download verified.	2026.10.15.9
8	Self-test after the server is ready	Pass	Self-test line with generation and prompt speed. Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.9 to .12
9	Resizable BAR off is detected and handled	Pass	Note with a How to turn it on link on the PC with Resizable BAR off.	2026.10.15.9
10	Closing the window keeps a background server running	Pass	Background mode keeps serving; a foreground run asks before stopping. Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.9, 2026.10.15.12
11	Reopening reattaches to the running server	Pass	Attach running names the engine: Running on HIP (AMD ROCm 10.0). Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.9, 2026.10.15.11, 2026.10.15.12
12	Sign out and back in	Pass	Standard Windows account, RX 7900 XT, background run attached. Signing out ended it, as documented (background mode runs only while signed in). After signing back in, FastLLM said the last background run ended without Stop; nothing was stuck and a new background start worked. On 2026.10.15.12 a background-ownership notice briefly showed after that start; 2026.10.15.13 fixes it.	2026.10.15.12, 2026.10.15.13
13	Stop	Pass	Foreground and background stop; after a background stop the window says Stopped. The background run ended. Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.11, 2026.10.15.12, 2026.10.15.13
14	Resizable BAR setting persists	Pass	—	2026.10.15.9
15	Uninstall offers to keep models and engines	Pass	Default No keeps them; the offer shows the real size (18,023 MB) and covers engine components.	2026.10.15.11
16	Layout	Pass	Long status lines end with an ellipsis and show the full text on hover. Also passed with a standard Windows account on the RX 7900 XT.	2026.10.15.11, 2026.10.15.12, 2026.10.15.13
17	Update banner for a newer release	Pass	FastLLM 2026.10.15.13 is available, with Release notes and Update (standard Windows account).	2026.10.15.11 to .12, 2026.10.15.12 to .13
18	Update from the window	Pass	Downloads, checks the Bitworks signature, runs the installer: FastLLM was updated to 2026.10.15.13 (standard Windows account).	2026.10.15.11 to .12, 2026.10.15.12 to .13
19	Reopen after the update	Pass	New version; no banner, no integrity warning.	2026.10.15.12, 2026.10.15.13
20	Check for updates setting persists	Pass	—	2026.10.15.9
21	Experimental community components switch	Pass	Explains experimental recipes; off by default.	2026.10.15.9
22/32	Engine components list and removal	Pass	Sizes shown; the runtime cannot be removed while the engine needs it.	2026.10.15.9, 2026.10.15.12
23/24	Choose another setup	Pass	Two setups for the card, both measured by Bitworks: Vulkan (built in) and HIP.	2026.10.15.9
27	Engine component consent window	Pass	Each component with its source and license links; nothing installs until every box is ticked. Names the model (Qwen3.8-27B UD-Q4_K_M). Wording and flow reviewed and approved by the Bitworks owner.	2026.10.15.9, 2026.10.15.12
28	License links	Pass	Each license link opens the right page: the GitHub license pages for llama.cpp b11658 and for the ROCm interop library at its pinned commit, and the AMD HIP SDK license page. Reviewed and approved by the Bitworks owner.	2026.10.15.9, 2026.10.15.12
35/36	No consent: FastLLM says so and starts Vulkan	Pass	Running on Vulkan (fallback: AMD license not accepted).	2026.10.15.9
37	A declined license is remembered	Pass	—	2026.10.15.9
38	HIP from Choose another setup	Pass	Components download and are verified; Running on HIP (AMD ROCm 10.0).	2026.10.15.9, 2026.10.15.12
39/40/41	Install engine components automatically (opt-in)	Pass	Confirmation with the license links; on: installs without asking; off: stays off.	2026.10.15.9
42	Background run names its engine	Pass	—	2026.10.15.9
43	Component install failure	Pass	The network could not be switched off on a remotely operated PC, so the install failure was forced locally (the components' staging folder was blocked). FastLLM said the engine components could not be installed, continued on Vulkan (Running on Vulkan (fallback: HIP components could not be installed)), and asked for consent again on the next start.	2026.10.15.12

Earlier engine tests

These tests ran on earlier FastLLM releases. They are kept for the record and describe the release named on each test, not the current download.

Tested with FastLLM 2026.10.15.13

Engine check

llama.cpp b11712 Windows Vulkan x64

Tested 2026-10-17 with FastLLM 2026.10.15.13. At that release: Bitworks lab comparison, not a FastLLM release; FastLLM still ships b10698 for Vulkan.

Method: Same model file, context, server arguments and environment as FastLLM's Vulkan setup for each card; arms in order shipped, candidate, shipped (drift control); FastLLM's own self-test (1 discarded warm-up + 2 measured runs; 512-token prompt; 128 generated tokens; server timings, median) plus a fixed correctness probe; standard Windows user.

Card	Run	Engine	Generation (tok/s)	Prompt (tok/s)	First token	Layers on GPU	Self-test	Correctness
AMD Radeon RX 7900 XT 20GB	Shipped, first run	llama.cpp b10698 Vulkan	34.1	538.3	954 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate	llama.cpp b11712 Vulkan	36.5	607.2	847 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Shipped, second run	llama.cpp b10698 Vulkan	33.6	539.1	951 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, first run	llama.cpp b10698 Vulkan	33.6	658.1	781 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate	llama.cpp b11712 Vulkan	34.5	682.2	752 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, second run	llama.cpp b10698 Vulkan	34.5	671.9	763 ms	66/66	Pass	Pass

Card	Generation vs shipped	Prompt vs shipped	Shipped-run drift (generation / prompt)
AMD Radeon RX 7900 XT 20GB	+7.9%	+12.7%	1.4% / 0.2%
AMD Radeon RX 7900 XTX 24GB	+1.4%	+2.5%	2.6% / 2.1%

Keep b10698 for now. The rule declared before the run needed a clear win on both cards: the RX 7900 XT improved clearly; the RX 7900 XTX did not regress but its gain is within run-to-run variation.

llama.cpp b11712 release on GitHub
SHA-256 7b80950ca99093f5a7b6d9ce2e2d52c5fa650f5fd3a583b1ad000dbd38ce6085

Adoption test

llama.cpp b11712 Windows Vulkan x64

Tested 2026-10-19 with FastLLM 2026.10.15.13. At that release: Bitworks lab test, not a FastLLM release; FastLLM still ships b10698 for Vulkan.

Method: Retest of the 2026-10-17 adoption test with heat controlled: five minutes idle before the first run and two minutes idle before every run. Same model file, context, server arguments and environment as FastLLM's Vulkan setup; runs interleaved shipped/candidate four times per card; FastLLM's own self-test on every run; on the first pair, FastLLM's API canary, its 3-case semantic smoke and its 24-case task-quality suite; standard Windows user. The rule is the same as on 2026-10-17.

Rule: Adopt only if every run passes, quality is not lower, no card is slower on average, and at least one card is clearly faster than every baseline run.

Thresholds checked by Bitworks: “no slower” means the candidate’s average is at least 98% of the shipped average for both generation and prompt; “clearly faster” means every candidate run beats every shipped run by at least 3% at generation, or by at least 5% at prompt processing.

Card	Run	Engine	Generation (tok/s)	Prompt (tok/s)	First token	Layers on GPU	Self-test	Correctness
AMD Radeon RX 7900 XT 20GB	Shipped, run 1	llama.cpp b10698 Vulkan	33.9	545.1	941 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 1	llama.cpp b11712 Vulkan	35.2	589.1	873 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Shipped, run 2	llama.cpp b10698 Vulkan	34.3	543.1	946 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 2	llama.cpp b11712 Vulkan	34.8	586.0	877 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Shipped, run 3	llama.cpp b10698 Vulkan	32.4	533.6	961 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 3	llama.cpp b11712 Vulkan	35.3	602.0	854 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Shipped, run 4	llama.cpp b10698 Vulkan	34.0	526.0	975 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 4	llama.cpp b11712 Vulkan	35.1	568.7	902 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 1	llama.cpp b10698 Vulkan	34.7	651.1	788 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 1	llama.cpp b11712 Vulkan	35.7	696.1	738 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 2	llama.cpp b10698 Vulkan	35.8	648.7	791 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 2	llama.cpp b11712 Vulkan	34.7	697.8	735 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 3	llama.cpp b10698 Vulkan	34.1	669.0	767 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 3	llama.cpp b11712 Vulkan	35.0	699.4	734 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 4	llama.cpp b10698 Vulkan	34.2	650.6	788 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 4	llama.cpp b11712 Vulkan	34.6	694.5	739 ms	66/66	Pass	Pass

Card	Check	Shipped (b10698)	Candidate (b11712)
AMD Radeon RX 7900 XT 20GB	API canary	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Semantic smoke	2/3	2/3
AMD Radeon RX 7900 XT 20GB	Task quality	21/24	21/24
AMD Radeon RX 7900 XTX 24GB	API canary	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Semantic smoke	2/3	2/3
AMD Radeon RX 7900 XTX 24GB	Task quality	21/24	21/24

Card	Mean generation vs shipped	Mean prompt vs shipped	No slower	Clearly faster than every shipped run
AMD Radeon RX 7900 XT 20GB	+4.3%	+9.2%	Yes	No
AMD Radeon RX 7900 XTX 24GB	+0.8%	+6.4%	Yes	No

Not adopted: FastLLM keeps b10698. With cool-downs between runs, quality was identical and no card was slower; b11712 was faster in every pair on the RX 7900 XT, and at prompt processing in every pair on the RX 7900 XTX, but not by the margin over every baseline run that the rule requires.

llama.cpp b11712 release on GitHub
SHA-256 7b80950ca99093f5a7b6d9ce2e2d52c5fa650f5fd3a583b1ad000dbd38ce6085

Earlier run: llama.cpp b11712 Windows Vulkan x64

An earlier run of the same adoption test; the newer run is above.

Tested 2026-10-17 with FastLLM 2026.10.15.13. At that release: Bitworks lab test, not a FastLLM release; FastLLM still ships b10698 for Vulkan.

Method: Follow-up to the engine check, aimed at adoption. Same model file, context, server arguments and environment as FastLLM's Vulkan setup; runs interleaved shipped/candidate three times per card; FastLLM's own self-test on every run; on the first pair, FastLLM's API canary, its 3-case semantic smoke and its 24-case task-quality suite; standard Windows user.

Rule: Adopt only if every run passes, quality is not lower, no card is slower on average, and at least one card is clearly faster than every baseline run.

Thresholds checked by Bitworks: “no slower” means the candidate’s average is at least 98% of the shipped average for both generation and prompt; “clearly faster” means every candidate run beats every shipped run by at least 3% at generation, or by at least 5% at prompt processing.

Card	Run	Engine	Generation (tok/s)	Prompt (tok/s)	First token	Layers on GPU	Self-test	Correctness
AMD Radeon RX 7900 XT 20GB	Shipped, run 1	llama.cpp b10698 Vulkan	34.3	538.7	952 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 1	llama.cpp b11712 Vulkan	34.5	592.7	868 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Shipped, run 2	llama.cpp b10698 Vulkan	33.0	522.4	983 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 2	llama.cpp b11712 Vulkan	34.0	582.9	879 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Shipped, run 3	llama.cpp b10698 Vulkan	32.2	511.3	1,004 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Candidate, run 3	llama.cpp b11712 Vulkan	34.8	550.9	932 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 1	llama.cpp b10698 Vulkan	34.1	642.0	800 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 1	llama.cpp b11712 Vulkan	33.7	664.7	773 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 2	llama.cpp b10698 Vulkan	34.3	670.2	767 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 2	llama.cpp b11712 Vulkan	35.2	701.2	732 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Shipped, run 3	llama.cpp b10698 Vulkan	32.9	659.9	778 ms	66/66	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Candidate, run 3	llama.cpp b11712 Vulkan	34.2	687.9	745 ms	66/66	Pass	Pass

Card	Check	Shipped (b10698)	Candidate (b11712)
AMD Radeon RX 7900 XT 20GB	API canary	Pass	Pass
AMD Radeon RX 7900 XT 20GB	Semantic smoke	2/3	2/3
AMD Radeon RX 7900 XT 20GB	Task quality	21/24	21/24
AMD Radeon RX 7900 XTX 24GB	API canary	Pass	Pass
AMD Radeon RX 7900 XTX 24GB	Semantic smoke	2/3	2/3
AMD Radeon RX 7900 XTX 24GB	Task quality	21/24	21/24

Card	Mean generation vs shipped	Mean prompt vs shipped	No slower	Clearly faster than every shipped run
AMD Radeon RX 7900 XT 20GB	+3.8%	+9.8%	Yes	No
AMD Radeon RX 7900 XTX 24GB	+1.7%	+4.1%	Yes	No

Not adopted: FastLLM keeps b10698. Quality was identical and no card was slower; b11712 was faster on average on both cards, but not beyond the spread of the baseline runs, which the rule required.

llama.cpp b11712 release on GitHub
SHA-256 7b80950ca99093f5a7b6d9ce2e2d52c5fa650f5fd3a583b1ad000dbd38ce6085

How we test and measure →

← Back to AMD Inference