Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

setup: accept the DeepSeek shard 1 with ISTA's 2026-10-26 chat template from files already here

Niko122194d7aab293a3
diff --git a/setup.py b/setup.py
index 3697b57e4..73471f68f 100644
--- a/setup.py
+++ b/setup.py
@@ -704,6 +704,14 @@ DSV41_ISTA_FILES = {
"DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00004-of-00004.gguf": (20527837312, "9b790a2297ed3cdede379fc984f4a52bb95f3b23724c797af11d1fa6a69db059"),
"DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf": (110595290816, "eb19f30d368d393784e1fc54fdfaffc427ebb2110bf3762fdf5a8fbb9a2d740a"),
}
+# ISTA re-published the first shard of each set on 2026-10-26 with a larger chat template (images in tool results; their PATCHING.md, and what their
+# patch_chat_template.py makes of a shard 1 already downloaded). setup downloads the pinned revision (HF_REVISIONS), so DSV41_ISTA_FILES stays the older
+# file; a newer shard 1 "already here" (--gguf-dir) is the same weights with the newer template and is accepted by its own size and SHA-256, never fetched.
+DSV41_NEW_SHARD1 = {
+ "DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ1_M-00001-of-00004.gguf": (39358841536, "a7be0d5a48e4d1c13c5535f9123d9131ab8964e5acf8a0c01d14970e4b7f4df9"),
+ "DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00004.gguf": (39535788064, "d9b949d55e27242b8afc7c2928749b888372d4c39682f9a131670d1c2555155e"),
+ "DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00004.gguf": (39075725728, "2dfe9711eac1f7814aa59154d11e8f27739f8b96875dc1e667eac9f8c756e1fe"),
+}
# The files the first release of these sets was downloaded as (helcig/tmp-oct26-2, 2026-10-23: model-2.00bpw-0000N-of-00004.gguf, model-2.25bpw-...,
# engram-IQ4_NL.gguf, usually in a 2.00bpw/ or 2.25bpw/ folder): testers and our boxes have them. They are the same weights: shards 2-4 are the
# same bytes as today's, and shard 1 and the Engram differ only in their GGUF header (metadata). setup accepts them from "files already here"
@@ -8340,7 +8348,12 @@ def install_dsv41(a, family, gpu, hip, amd, cuda_tk, multi, port, ram, adopted,
download((md.get("hf") or fam["hf"]) + f.name, f)
if f.name in DSV41_LEGACY_FILES:
ok(f"{f.name}: a file of the first release of this set (the same weights); it is used as it is")
- verify_sha256(f, *fam["sha256"][f.name], own=bool(a.gguf_dir) or f.name in DSV41_LEGACY_FILES)
+ pin = fam["sha256"][f.name]
+ new = DSV41_NEW_SHARD1.get(f.name)
+ if new and a.gguf_dir and f.is_file() and f.stat().st_size == new[0] and f.stat().st_size != pin[0]: # shard 1 with the 2026-10-26 chat template
+ pin = new
+ ok(f"{f.name}: the version with ISTA's newer chat template (the same weights); it is used as it is")
+ verify_sha256(f, *pin, own=bool(a.gguf_dir) or f.name in DSV41_LEGACY_FILES)
for f in shared: # the Engram, the drafter, the metadata shard: once for every official set
if f.exists() and (done(f) or own_files[f]):
ok(f"{f.name} already here ({f.parent})")
diff --git a/tools/test_setup_dsv41.py b/tools/test_setup_dsv41.py
index 376789b26..c6470140c 100644
--- a/tools/test_setup_dsv41.py
+++ b/tools/test_setup_dsv41.py
@@ -575,6 +575,10 @@ class Smallest(unittest.TestCase):
self.assertEqual(setup.DSV41_ISTA_FILES[d["shards"][3]][0], 2188841824)
self.assertEqual(sum(setup.DSV41_ISTA_FILES[n][0] for n in d["shards"]), 120_721_120_256)
self.assertTrue(all(n not in setup.DSV41_LEGACY_NAMES.values() for n in d["shards"])) # no first-release names for it
+ for n, (size, sha) in setup.DSV41_NEW_SHARD1.items(): # shard 1 after ISTA's 2026-10-26 template update
+ self.assertIn(n, setup.DSV41_ISTA_FILES)
+ self.assertNotEqual(setup.DSV41_ISTA_FILES[n][0], size)
+ self.assertNotEqual(setup.DSV41_ISTA_FILES[n][1], sha)
cmd = setup.dsv41_pack_cmd(Path("/m"), d["shards"], Path("/s/e.gguf"), Path("/s/m.gguf"), Path("/s/d.gguf"), Path("/p"))
self.assertIn("--dspark-noise-token", cmd)
self.assertNotIn("--skip-tensors", cmd)