The roster, as of 26 September 2026

Twenty-three endpoints on a one-model machine

The machine is one desktop: an NVIDIA GeForce RTX 5090 with 32,607 MiB of graphics memory, an Intel Core Ultra 9 285K and 188 GiB of system RAM. Row 1 also uses a second machine, an ASUS ROG Flow Z13 with 128 GB of unified memory, joined to the desktop by a direct Thunderbolt cable; its cells say so. Every model is served by llama-server from llama.cpp, and every start script refuses to launch while another model is up; the package holds one recorded test in which seven separate guards each refused a launch.

Changes since 15 September, as our operator records have them: Ornith-1.5-35B-A3B and GLM-5.3-Flash were installed on 16 September and MiniMax M2.7 on 21 September. Kimi K3 was retired on 15 September and Kimi K2.7-Code's weights were removed on 17 September. MiniMax M3 serves a different file since 21 September, Laguna S 2.1 a 262,144-token window since the same day, and Qwen3.6-27B a different file since 26 September; our records also give it a 262,144-token window from that day, which the runs on this page loaded by passing it explicitly. Qwen3.8-27B was installed at a 262,144-token window on the afternoon of 26 September; its start script still accepts 131,072. Qwen3.5-122B-A10B and Ornith-1.5-35B-A3B were installed at 262,144 the same afternoon, MiniMax M2.7 at 196,608, its native ceiling, that evening, and GLM-5.3-Flash at 262,144 later that evening; a start of each of these four with no window passed was served that window. Each row's window cell says whether a run that passed no window shows the installed default. A count of twenty-three installed endpoints is not a capacity claim: one runs at a time.

What we ran, exactly

The figures come from several exercises between 12 and 26 September 2026. Two methods never share a column: a request sent straight to the server (direct), and the same kind of request made through an agent framework.

Why there are two speed columns and not one

A model measured directly and through a framework does not give the same number. Gemma-4-26B read at 8,972.3 tokens a second directly at 47,983 tokens and at an estimated 6,726.8 through the framework at 53,209. Qwen3-235B spoke at 9.3 through the framework on a 62-token answer at 53,458 tokens, and at 5.51 directly on a 478-token letter at 20,063. Neither figure is wrong: they answer different questions, and mixing them in one column would be the worst error available on this page. Every speaking figure below also says what it timed, because a short answer and a page of prose are not the same measurement.

The twenty-three, with their newest figures

Every speed below is measured, on the dates given, one request at a time. Files and byte counts are read from the files on this machine (a listing taken on 26 September); maker names are the makers' own. "Short answer" means the model's reply to a quote-the-code question, and its token count includes hidden reasoning. Reading speeds after the batch-size sweep name the flag and the depth. Row numbers match ROSTER_23.csv in the data package.

"Speaking" is generation speed, tokens produced per second; "reading" is prompt processing, tokens of prompt consumed per second, which decides how long you wait before the first word. Every figure's own qualifier (file, window, setting, date) controls. Rows without a date in a cell carry the date in the last column. Every row that shows a planted-code read quoted back all the codes at the depths listed; row 13 is the only failed tool call at depth. The framework's reading figures are its own estimates and its speaking figures are short answers, as section 02 says.

Row notes

What a batch setting did, and did not, change

The largest movements in this table are reading speeds on models whose experts sit in system memory. Each read is about 48,000 tokens, the same prompt at both settings, except GLM-5.3-Flash's default-setting read of 24,008 tokens; all six at a 131,072-token window, measured on 21 September through each model's own start script (Qwen3.5-122B, Ornith and GLM-5.3-Flash have served 262,144 since 26 September; rows 6, 8 and 9):

All measured, 2026-09-21; in these six, the planted code came back in the answer at every setting that loaded. The speaking figures in those runs moved by 6.2 percent at most, Qwen3.5-122B's 30.81 to 32.72 (arithmetic on the same files). At 262,144 the start scripts of Qwen3.5-397B, Qwen3.5-122B and Mistral Small 4 keep -b 2048 -ub 512: the larger buffer does not fit there; since 26 September Ornith's has used -b 2048 -ub 2048 there and GLM-5.3-Flash's -b 4096 -ub 1024. The models that fit whole on the card kept the defaults: Qwen3.8-27B, GLM-4.7-Flash, the two Gemmas and Muse Glimmer. On each one's best rung, reading moved from a 0.6 percent loss (Muse Glimmer) to a 34.5 percent gain (Gemma-4-26B-A4B at -ub 2048, which spoke 6 percent slower), arithmetic on each model's rungs. Qwen3.8-27B's rungs were run at 131,072, its installed window on 21 September. Qwen3.6-27B kept them too; its rungs were run on its earlier Q4 file at 131,072. GLM-4.7-Flash at -ub 2048 put the code only in its hidden reasoning and returned an empty answer.

Three rows moved for other reasons: MiniMax M2.7 and Laguna S 2.1 had inherited -ub 128, Laguna with -b 512 (row notes 18 and 20), and MiniMax M3 changed file. DeepSeek's two services moved further still, and the two-machine study sets them side by side. How to run the sweep on your own models, and the one value that passed every read and still crashed a server on a specific input, are on the batch-size study.

What changed since 15 September

This list was first compiled on 15 September. The later runs corrected figures in our own summary table and in that first list. We publish those corrections the same way we publish everything else.

What held up, in the same breath: every read behind this page's figures that asked for a planted code returned it in its answer, except GLM-4.7-Flash at -ub 2048, whose code stayed in its hidden reasoning (section 04). The letter runs' edit-and-reread requests, capped at 40 tokens of reply, did not return the code at any depth (row notes 5, 7 and 10).

What this does not show

• No measure of quality. Speed, window, recall of a planted code and whether a tool call came back are the only things measured. A model can be the fastest row here and the wrong choice for your work.
• Short answers are not prose. Where a row's only speaking figure at depth is a short answer, it says so; eight rows have prose at depth (1, 5, 6, 7, 8, 9, 10 and 18), and most have a paragraph at a short prompt. A thinking model's short answer counts its reasoning.
• Figures from different exercises are not a ranking. Rows were measured on different days, at different depths and settings, and each cell says which. Compare within a column and a date.
• One desktop. These are honest samples on one machine, not benchmarks, and not comparable with a vendor's published throughput or with a hosted service.
• Windows above the served one are not claimed here. "On request" means the start script offers a larger window and a September run loaded it: Qwen3.5-397B and Mistral Small 4 read about 231,000 tokens at 262,144 (15 September, at -b 2048 -ub 512, the setting their scripts keep there), Qwen3.8-Flash-Next read 344,981 and 459,911 tokens at its two larger windows (20 September, at llama.cpp's default batch sizes, before its batch change), and MiniMax M3 58,307 at 196,608 (20 September, -ub 512). What a full window costs is on its own page.
• The one-at-a-time constraint is this machine's. On more hardware several of these would run side by side.
• Two rows have no figure for what they serve. That is a statement about our September, not about the models.

Related pages on this site

• Method: the labels used on this page.
• The batch-size study: the sweep behind most reading figures here, and how to tune your own.
• Two boxes, one model: rows 1 and 3, and the models run across two machines.
• What a full window costs: the long reads behind the window column.
• DeepSeek V4 Flash, Inkling-Small, Qwen3.8-27B, Qwen3.8-27B at 256K, the Qwen3.5 pair, Mistral Medium 3.5, Muse Glimmer 30B and Kimi K3: field cards for models on or formerly on this list, each with its own dates.
• Which Qwen should I run?: the family guide behind seven of the rows.
• Empty Answer: why a thinking model's short answer can come back empty, which is why the counts here include its reasoning.

Sources and artifacts

The data package holds the redacted primary files these figures come from, a few clearly identified extracts, one rebuilt machine-readable table, and the vendor documents cited at pinned revisions. Machine paths, addresses, port numbers, service names and our internal keys for each model are removed; so is a column naming private work. Nothing that carries a measurement is altered, and no figure is recomputed except where a file says it is a derivation.

• data/README.md: every file, where it came from, what was removed from it, and the places where the package looks like it argues with itself.
• data/NUMBERS.md: every figure on this page mapped to its file and field.
• data/ROSTER_23.csv: the table, machine readable, with the file and date behind each figure.
• The runs of 19 to 26 September: the batch-size sweep's result files and logs, the long reads, the paragraph replies, the letters at depth, the MiniMax records, the new framework record, and the 26 September runs of Qwen3.8-Flash-Next, Qwen3.8-27B, Qwen3.5-122B-A10B and Ornith-1.5-35B-A3B and GLM-5.3-Flash at 262,144 and of MiniMax M2.7 at 196,608, and that evening's starts with no window passed.
• The runs of 12 to 15 September: the direct sweep, the framework check's ledgers and per-model records, the refusal-guard test, and what the model files say.

If a number on this page disagrees with a file in its package, the file is right and the page is wrong. Tell us at hello@graphometer.ai and the correction goes on the page with its date.

This page is dated on purpose. It is a snapshot as of 26 September 2026, not a living list; a later roster will be a later page with a later date.