Flint · file← Flint

Flint

The only file: Recursive's own Q4_1, holding the 4-bit weights the model was trained with. Tested in full once, one conversation at a time with 64K of context.

Size: 26.0 GB
Status: Benched 24 Oct 2026
Scored runs: 15
Smallest Mac: 36 GB measured, 64K context

Results15 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
chainwalkflint-q4_1-np1-c64k-b11846	50.0%	9.5–90.5	Passed 1/2	llama.cppextraction	2026-10-24
contractflint-q4_1-np1-c64k-b11846	60.4%	46.3–73.0	Passed 29/48	llama.cppextraction	2026-10-24
enrichmentflint-q4_1-np1-c64k-b11846	77.8%	65.1–86.8	Passed 42/54	llama.cppextraction	2026-10-24
epistemicsflint-q4_1-np1-c64k-b11846	95.8%	86.0–98.8	Passed 46/48	llama.cppextraction	2026-10-24
groundingflint-q4_1-np1-c64k-b11846	77.6%	69.5–84.0	Passed 97/125	llama.cppextraction	2026-10-24
needleflint-q4_1-np1-c64k-b11846	60.0%	50.7–68.7	Passed 66/110	llama.cppextraction	2026-10-24
nullschemaflint-q4_1-np1-c64k-b11846	70.8%	59.5–80.1	Passed 51/72	llama.cppextraction	2026-10-24
screeningflint-q4_1-np1-c64k-b11846	98.3%	91.1–99.7	Passed 59/60	llama.cppextraction	2026-10-24
synthesisflint-q4_1-np1-c64k-b11846	78.3%	66.4–86.9	Passed 47/60	llama.cppextraction	2026-10-24
taggerflint-q4_1-np1-c64k-b11846	82.1%	76.7–86.4	Passed 197/240	llama.cppextraction	2026-10-24
tagger_nullableflint-q4_1-np1-c64k-b11846	60.8%	54.5–66.8	Passed 146/240	llama.cppextraction	2026-10-24
toolcallflint-q4_1-np1-c64k-b11846	55.0%	42.5–66.9	Passed 33/60	llama.cppextraction	2026-10-24
toolsflint-q4_1-np1-c64k-b11846	100.0%	79.6–100.0	Passed 15/15	llama.cppextraction	2026-10-24
toolscaleflint-q4_1-np1-c64k-b11846	95.0%	83.5–98.6	Passed 38/40	llama.cppextraction	2026-10-24
worldviewflint-q4_1-np1-c64k-b11846	75.0%	59.8–85.8	Passed 30/40	llama.cppextraction	2026-10-24

Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.

SetupHow it was run

flint-q4_1-np1-c64k-b11846

Of record

The measurement of record for this model.

The run’s own note: Single slot at 64K (benchmark rule). shapegen and sitebuild excluded: unfinished suite / harness dataset fault.

Server: 1 slot, 64K of context per request, KV cache f16 (default), flash attention on, 1 suite at a time
Runtime: llama.cpp b11846 (1a24df163), plain upstream
Speed: decode 74.2 tokens/s, prefill 345.2 tokens/s (444.7 on prompts over 16K), medians. Single stream, fleet resident except the vacated play seat (contended).
Peak memory: 26.4 GB resident, with 64K of context allocated
Weights: 26.0 GB file · sha256 43c07baccdba42eb…
Coherence check: passed
Left out by plan: shapegen, sitebuild
Machine: Mac Studio (M1 Max, 64 GB)
Released: 24 October 2026, bundle 2026-10-24-746b3da7

Memory by context

The same file and server, started at each context size in turn: what it held in memory and how fast it wrote on a short prompt. It measures memory and serving, not how well the model uses a long context. Released 24 October 2026, bundle 2026-10-24-51f2fa68.

The ladder’s own note: Context ladder 8K to native 256K, single slot, f16 KV, play seat vacated, other seats up. Decode on a trivial prompt per rung; measures allocation and serving, not long-context quality. First rung sampled before mmap'd weights were fully resident.

Context	Resident memory	Swap in use, whole Mac	Decode	Load
8K	21.4 GB	13.43 GB	77.2 tokens/s	15.1 s
16K	24.86 GB	13.43 GB	76.8 tokens/s	3 s
32K	25.3 GB	13.43 GB	77.1 tokens/s	3 s
64K	26.18 GB	13.43 GB	77.3 tokens/s	3 s
128K	27.93 GB	13.43 GB	77.1 tokens/s	3 s
256K	31.44 GB	13.43 GB	77.3 tokens/s	3 s

flint-q4_1-np4-b11846

Stopped

Stopped part-way. Its partial suites are in the run log and aren't scored.

The run’s own note: Stopped after about 5 minutes: 16K per request is below the 64K Hermes needs, so the single-slot rule replaced it. Partial suites only; not scored as results.

Server: 4 slots, 16K of context per request, KV cache q4_0, flash attention on, 4 suites at a time
Runtime: llama.cpp b11846 (1a24df163), plain upstream
Speed: not recorded
Peak memory: 24.3 GB resident, with 64K of context allocated
Weights: 26.0 GB file · sha256 43c07baccdba42eb…
Coherence check: not recorded
Left out by plan: none
Machine: Mac Studio (M1 Max, 64 GB)
Released: 24 October 2026, bundle 2026-10-24-80653652

The recordHow it was made · how it was served · what went wrong · what happened to it

Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.

File	`flint-int4-q4_1.gguf`
Format	GGUF Q4_1 (trained INT4)
Where from	downloaded
Status	benched 2026-10-24
Hugging Face	RecursiveSI/flint-int4-gguf

How it was made

Downloaded, not made here.

field	value
repo	`RecursiveSI/flint-int4-gguf`
file	`flint-int4-q4_1.gguf`
revision	`9c47151036bc2c34331142dda04488bceab2bc9f`
size	25,992,966,976 bytes (`ls -l`, matches the HF listing)
sha256	`43c07baccdba42eb7528a9630ee15699b1942c297291dc1b39a061f508ee1a4b` (`shasum -a 256`, equals the HF LFS oid)
downloaded	2026-10-23 21:20–21:37Z, `huggingface_hub` 1.33.0 `hf_hub_download` with `HF_XET_HIGH_PERFORMANCE=1`; ~320 Mbps on en0
local path	`models/flint-40b-a2.5b/flint-int4-q4_1.gguf` (with the repo's README.md and LICENSE)
asked by	Alan: "can we download flint?"

How it was served

Machine: a Mac Studio (Mac13,1): Apple M1 Max, 64 GB.

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / --jinja / --reasoning-format	profile	why this label exists
`flint-q4_1-np4-b11846`	2026-10-23 22:30–22:35Z	`llama.cpp-b11846/build/bin/llama-server`, build 11846 (`1a24df163`)	`-c 65536 -np 4` (16,384 per slot) / q4_0 K and V / `-fa on` / `--jinja` / default / `-ngl 999 --cache-ram 0`	extraction	the first pre-registered plan. Stopped after ~5 minutes, not scored: 16K per request is below Hermes's 64K (family decision log, 22:30–22:35Z)
`flint-q4_1-np1-c64k-b11846`	2026-10-23 23:46Z – 2026-10-24 02:02Z	`llama.cpp-b11846/build/bin/llama-server`, version 0.7.0-dev, build 11846 (`1a24df163`)	`-c 65536 -np 1` (server log: `n_slots = 1, n_ctx_slot = 65536`) / KV default f16 / `-fa on` / `--jinja` / default / `-ngl 999 --cache-ram 0`	extraction	the measurement of record. Alan's single-slot 64K rule, run by a one-off script through `eval/swapseat_gauntlet.sh` with `VACATE=1` and `PAR=1`, 15 suites
`flint-q4_1-np1-b11846` (context ladder)	2026-10-24 09:09–09:10Z	the same binary	`-c 8192 … 262144 -np 1 -fa on --jinja -ngl 99`, KV default f16	—	`eval/ctx_ladder.py` with `LADDER_BIN`. The play seat was vacated (Alan: "Both, vacate for Flint"), Ornith restored 09:10:43Z

Seats held: none.

Decisions

2026-10-23: downloaded as the family's only variant, because it is the native trained-INT4 format.
2026-10-23: Alan approved vacating the play seat (Ornith 35B, 23 GB) for this run. The box was 13.4 GB into swap with all seats up.
2026-10-24: Alan approved the context ladders with the play seat vacated for Flint ("Both, vacate for Flint").

What went wrong

The np4 run was stopped, not scored. flint-q4_1-np4-b11846 ran 22:30–22:35Z at 16K per request, then stopped under Alan's single-slot 64K rule. Its partial result files (grounding 56, tagger 17, tagger_nullable 20, needle 5 rows) stay under that label, unscored, and are marked incomplete in bundle 2026-10-24-80653652.
The run log's llama_build and server_flags will be wrong for both labels when the run log is backfilled. The run-log script reads a fixed binary path and the router. The truth is build 11846 with the flags under How it was served. LFM2.5-24B-A2B's runs have the same problem.
The ladder's KV column reads "?": b11846's server log no longer matches ctx_ladder.py's KV-size regex. KV is taken from RSS growth instead.

What happened to it

Downloaded 2026-10-23 → gauntlet complete 2026-10-24 (single slot, 64K) → context ladder 2026-10-24 (to native 256K, 31.4 GB) → decide/1 and edgefill pending → seat decision pending (Alan). Re-downloadable from the repo and revision above.