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++

Catalog: Gyro-S and Gyro-M (agentionai/Qwen3.8-Flash-Next-Gyro-GGUF) in setup, the Terminal, the app and the MCP server

Niko1221ba64156fd43a
diff --git a/apps/host/src/lib/native-catalog.ts b/apps/host/src/lib/native-catalog.ts
index 0813f75be..aee650b47 100644
--- a/apps/host/src/lib/native-catalog.ts
+++ b/apps/host/src/lib/native-catalog.ts
@@ -2,16 +2,16 @@ import type {NativeFamily,NativeSize} from './native';
import {tr} from './i18n';
export interface CatalogVersion extends NativeSize{sourceFamily:string;edition:string;images:boolean}
export interface CatalogModel extends Omit<NativeFamily,'sizes'>{sizes:CatalogVersion[]}
-export function nativeVersionLabel(model:string){return tr(({Q2_0:'Fastest',IQ2_XS:'Compact',IQ3_XXS:'Balanced',IQ3_S:'High quality',IQ1_M:'Coding'} as Record<string,string>)[model]??model);}
+export function nativeVersionLabel(model:string){return tr(({Q2_0:'Fastest',IQ2_XS:'Compact',IQ3_XXS:'Balanced',IQ3_S:'High quality',IQ1_M:'Coding','Gyro-S':'Smallest','Gyro-M':'Small memory'} as Record<string,string>)[model]??model);}
export const catalogVersionLabel=(version:CatalogVersion)=>{const label=nativeVersionLabel(version.model);return [tr(version.edition),label===version.model&&version.edition?'':label].filter(Boolean).join(' · ');};
/** Display groups are distinct from the engine's installation family IDs. */
export function nativeCatalogModels(families:NativeFamily[]):CatalogModel[]{
if(!families.length)return[];
const groups:CatalogModel[]=[{family:'qwen3.8',title:'Qwen3.8-Flash-Next',about:'',experimental:false,images:false,sizes:[]},{family:'qwen3.6',title:'Qwen3.6-35B-A3B',about:'',experimental:false,images:false,sizes:[]}];
for(const family of families){
- const target=/qwen3[._-]?6|35b[-_]?a3b/i.test(family.title+' '+family.family)?groups[1]:['qwen','swift','coder','unsloth'].includes(family.family)||/qwen3[._-]?8|flash.next/i.test(family.title)?groups[0]:undefined;
+ const target=/qwen3[._-]?6|35b[-_]?a3b/i.test(family.title+' '+family.family)?groups[1]:['qwen','swift','coder','unsloth','gyro'].includes(family.family)||/qwen3[._-]?8|flash.next/i.test(family.title)?groups[0]:undefined;
if(!target)continue;
- const edition=target===groups[0]?({swift:'Swift 1.5',coder:'Coder',unsloth:'Unsloth'} as Record<string,string>)[family.family]??'':'';
+ const edition=target===groups[0]?({swift:'Swift 1.5',coder:'Coder',unsloth:'Unsloth',gyro:'Gyro'} as Record<string,string>)[family.family]??'':'';
target.images ||= family.images;
target.sizes.push(...family.sizes.map(size=>({...size,sourceFamily:family.family,edition,images:family.images})));
}
diff --git a/setup.py b/setup.py
index c9b5fe0d7..811921f99 100644
--- a/setup.py
+++ b/setup.py
@@ -71,6 +71,7 @@ HF_REVISIONS = {
"ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF": "5348543e0147355ac9cbcb031184a3546350988e", # 2026-09-29
"unsloth/Qwen3.8-Flash-Next-GGUF": "38bb39ee97821de2c9009abb7e93950eec396e66", # 2026-09-30
"unsloth/Qwen3.6-35B-A3B-MTP-GGUF": "5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", # 2026-10-12
+ "agentionai/Qwen3.8-Flash-Next-Gyro-GGUF": "1ada08422e404aca51ab9df665f089f34c419c6b", # 2026-10-13
}
@@ -244,6 +245,18 @@ NVFP4_REPOS = {
# the first release engine with NVFP4 experts (a source build of this checkout has them whatever its number says)
NVFP4_ENGINE = (0, 1, 44)
+# agentionai's Gyro files (Qwen3.8-Flash-Next with trellis-coded, Hadamard-rotated experts, TQ2_T / TQK6 / TQK7): one
+# GGUF per size, the n-gram table inside it (hence the 58 / 92 GB downloads for 27.6 / 35 GB of memory). The engine
+# reads them from 0.1.46 on (ported from agentionai/Strata rc2, MIT); the sizes and SHA-256 are the Hub's LFS pointers
+# of the revision pinned in HF_REVISIONS. The model's license is Qwen's.
+GYRO_ENGINE = (0, 1, 46)
+GYRO_FILES = {
+ "Qwen3.8-Flash-Next-Gyro-S-TQ1_0.gguf":
+ (58487015808, "c6ea1b08ca312f99095b27588ba35976b53f948268e7f91e7761926aa01d9c87"),
+ "Qwen3.8-Flash-Next-Gyro-M-TQ2_0.gguf":
+ (91965774272, "7aa36d15131d255179ab6c87fa9c9f81a0b541fc9e81e959d4c3e29eb20feb5f"),
+}
+
MODELS = {
# the original model only for now: Swift 1.5's Q2_0 files split one layer's experts across the two shards, which
# the pack tool (tools/iq_pack.py) cannot prepare yet (#171)
@@ -298,6 +311,17 @@ MODELS = {
# (the HIP and SYCL engines have no NVFP4 kernels). Images: both packs keep Qwen's vision tower unquantized, so the
# original's image encoder (ISTA's mmproj, the same file as for every Flash-Next model) reads pictures for them;
# checked end to end on an RTX 5070 and an RTX 3060 against IQ3_XXS's answers on the same picture.
+ # Gyro (agentionai): "ram_gb" is the memory the weights need in total, VRAM and RAM together ("mem_total"): Gyro-S
+ # 27.6 GB, Gyro-M 35 GB (the rest of the download is the n-gram table, read from the SSD). Neither is anyone's
+ # default: setup and the MCP server mention them where the other sizes do not fit. No images yet (untested).
+ "Gyro-S": {"about": "the smallest option, for PCs with little memory (27.6 GB of VRAM + RAM together); lower quality "
+ "than the 2-3-bit sizes, and the 2-bit Q2_0 decodes faster",
+ "download_gb": 58.5, "ram_gb": 27.6, "arena_gb": 25.8, "families": ("gyro",), "mem_total": True,
+ "file": "Qwen3.8-Flash-Next-Gyro-S-TQ1_0.gguf", "shards": 1, "engine": GYRO_ENGINE, "vision": False},
+ "Gyro-M": {"about": "about IQ3_XXS quality with about 20% less GPU memory (35 GB of VRAM + RAM together); decodes "
+ "about 20% slower than Q2_0 on 12 GB cards",
+ "download_gb": 92.0, "ram_gb": 35.0, "arena_gb": 32.1, "families": ("gyro",), "mem_total": True,
+ "file": "Qwen3.8-Flash-Next-Gyro-M-TQ2_0.gguf", "shards": 1, "engine": GYRO_ENGINE, "vision": False},
"NVFP4": {"about": "4-bit NVFP4 experts, GPTQ (sergqwer's strata-nvfp4 pack); RTX 50 has FP4 tensor cores, other "
"NVIDIA cards run it at 8-bit precision", "download_gb": 128.0, "ram_gb": 80, "arena_gb": 68.0,
"families": ("nvfp4",), "nvidia_only": True, "vision": True, "repo": NVFP4_GPTQ_REPO,
@@ -420,6 +444,16 @@ FAMILIES = {
"sha256": {**UNSLOTH_SHARDS, **UNSLOTH_IQ4_XS_SHARDS, **UNSLOTH_BIG_SHARDS}},
# MODEL-UPDATE Phase 4: the original model with NVFP4 experts, as ready-made packs (NVFP4_REPOS): the files come
# as the engine reads them (no GGUF to pack), each size has its own repository ("repo" in MODELS)
+ # agentionai's Gyro quantizations (see GYRO_FILES): after the regular families, never the default
+ "gyro": {"title": "Qwen3.8-Flash-Next (Gyro)", "by": "Gyro quantizations by agentionai",
+ "about": "Gyro-S (27.6 GB of memory) and Gyro-M (35 GB): the smallest sizes, for PCs with little memory; "
+ "58 and 92 GB downloads (the files include the n-gram table)",
+ "hf": hf("agentionai/Qwen3.8-Flash-Next-Gyro-GGUF"), "file": "Qwen3.8-Flash-Next-{q}.gguf", "shards": 1,
+ "tag": "gyro-", "name": "qwen3.8-flash-next-gyro", "vision": False, "pack_args": ["--compat-bf16"],
+ "mmproj_hf": hf("ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF"), "mmproj": "mmproj-Qwen3.8-Flash-Next-BF16.gguf",
+ "sha256": GYRO_FILES,
+ "license": "Qwen Community License 1.0 (Qwen3.8-Flash-Next's; the Gyro files are agentionai's quantization): "
+ "https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF"},
"nvfp4": {"title": "Qwen3.8-Flash-Next (NVFP4)", "by": "NVFP4 packs by sergqwer (strata-nvfp4)",
"about": "4-bit NVFP4 experts, NVIDIA cards; a 128 GB download, ~80 GB of RAM to keep every expert in RAM",
"tag": "nvfp4-", "name": "qwen3.8-flash-next-nvfp4", "ready": True,
@@ -2290,14 +2324,20 @@ STRIX_HALO_MIN_GB = 80 # UD-IQ4_XS keeps all of its experts in memory fr
RECOMMENDED_SIZES = {"qwen": ("IQ3_S",), "swift": ("IQ3_S",), "unsloth": ("UD-Q4_K_XL", "UD-IQ4_XS")} # owner's pick, 0.1.44
-def recommended_size(family: str, names: list, ram: float) -> str:
+def mem_have(model: str, ram: float, vram: float = 0.0) -> float:
+ """The memory this PC offers a model, in the unit of its MODELS "ram_gb": RAM, or for a "mem_total" model (Gyro) the
+ VRAM and the RAM together."""
+ return ram + vram if MODELS[model].get("mem_total") else ram
+
+
+def recommended_size(family: str, names: list, ram: float, vram: float = 0.0) -> str:
"""The size the menu pre-selects (and --yes takes): the family's recommended size when this PC's RAM fits it
(UD-Q4_K_XL before UD-IQ4_XS), else the largest size that fits, else the first. Recommend, never force: every
other size stays a choice."""
for m in RECOMMENDED_SIZES.get(family, ()):
- if m in names and ram >= MODELS[m]["ram_gb"]:
+ if m in names and mem_have(m, ram, vram) >= MODELS[m]["ram_gb"]:
return m
- fitting = [m for m in names if ram >= MODELS[m]["ram_gb"]]
+ fitting = [m for m in names if mem_have(m, ram, vram) >= MODELS[m]["ram_gb"]]
if fitting:
return max(fitting, key=lambda m: MODELS[m]["download_gb"])
return "UD-IQ4_XS" if family == "unsloth" and "UD-IQ4_XS" in names else names[0]
@@ -2378,6 +2418,9 @@ def family_menu_note(family: str, hip: bool, gpu, avx2: bool = True, avx512: boo
"expert in RAM" + (f"; {cpu}" if cpu else "") + "]")
if family == "unsloth" and hip:
return " [UD-IQ4_XS is the 4-bit choice on AMD and Intel cards; NVFP4 runs on NVIDIA only]"
+ if family == "gyro":
+ return (" [for PCs with little memory: VRAM + RAM together 27.6 GB (Gyro-S) or 35 GB (Gyro-M)"
+ + ("; AMD is experimental, not for Intel Arc" if hip else "") + "]")
return ""
@@ -6351,7 +6394,7 @@ def main() -> int:
+ ": Strata will run, but most experts stay on the CPU and it will be slow")
ram = ram_gb()
cpu, avx2, avx512 = cpu_info()
- need = min(d["ram_gb"] for d in MODELS.values())
+ need = min(d["ram_gb"] for d in MODELS.values() if not d.get("mem_total")) # Gyro counts VRAM + RAM: not a RAM floor
low_ok = low_ram_fits("IQ1_M", ram, low_ram_vram(gpu)) and a.low_ram != "off" # the smallest model, mapped
if ram < need - 4 and not a.check and not low_ok:
# every model keeps ALL its experts in RAM (23+ GB); VRAM only holds a copy of the most-used ones, so a
@@ -6397,7 +6440,10 @@ def main() -> int:
say()
any_fits = False
for m, d in MODELS.items():
- verdict = "fits" if ram >= d["ram_gb"] else "tight" if ram >= d["ram_gb"] - 8 else "does not fit"
+ have_gb = mem_have(m, ram, gpu["vram_gb"])
+ verdict = "fits" if have_gb >= d["ram_gb"] else "tight" if have_gb >= d["ram_gb"] - 8 else "does not fit"
+ if d.get("mem_total"): # Gyro: VRAM + RAM together; the rest of the download is read from the SSD
+ verdict += f" ({ram:.0f} GB RAM + {gpu['vram_gb']:.0f} GB VRAM)"
if d.get("budget") and not unsloth_budget(m, ram, gpu): # Phase 2: all of them in RAM
verdict = ("EXPERIMENTAL, " if d.get("experimental") else "") + "fits (all of its experts in RAM)"
elif d.get("budget"):
@@ -6416,8 +6462,8 @@ def main() -> int:
if hip and d.get("repo"): # the ready-made NVFP4 packs: no AMD kernels at all
verdict = "NVIDIA only (the AMD engine has no NVFP4 kernels)"
else:
- any_fits = any_fits or not verdict.startswith("does not fit")
- say(f" {m:8s} needs ~{d['ram_gb']} GB RAM: {verdict}")
+ any_fits = any_fits or (not verdict.startswith("does not fit") and not d.get("mem_total"))
+ say(f" {m:8s} needs ~{d['ram_gb']} GB {'of VRAM + RAM' if d.get('mem_total') else 'RAM'}: {verdict}")
if not any_fits:
# #977: every size says "does not fit", so the verdict says so too (and the exit code, for scripts). It only
# reports: --model NAME --yes still installs one anyway.
@@ -6474,6 +6520,13 @@ def main() -> int:
for i, m in enumerate(names, 1):
d = MODELS[m]
fit = "" if ram >= d["ram_gb"] else f" <- needs {d['ram_gb']} GB RAM, you have {ram:.0f}"
+ if d.get("mem_total"): # Gyro: the download includes the n-gram table, the memory figure does not
+ have_gb = mem_have(m, ram, gpu["vram_gb"])
+ fit = "" if have_gb >= d["ram_gb"] else \
+ f" <- needs {d['ram_gb']} GB of VRAM + RAM together, you have {have_gb:.0f}"
+ say(f" {i}) {m}{size_tag(family, m, installed)} {d['about']}; download {d['download_gb']:.1f} GB "
+ f"(the n-gram table is in the file), memory {d['ram_gb']} GB{fit}")
+ continue
if d.get("budget"):
keeps = (f"~{resident_budget_gib(m, ram)} GB" if unsloth_budget(m, ram, gpu, a.resident_budget_gib)
else "all")
@@ -6484,7 +6537,7 @@ def main() -> int:
fit = (f" <- fits in the low-RAM mode (the GPU holds ~{100 * low_ram_gpu_share(m, low_ram_vram(gpu)):.0f}%, "
+ ("the rest in RAM)" if low_ram_resident(m, ram, low_ram_vram(gpu)) else "the rest from the SSD)"))
say(f" {i}) {m:8s}{size_tag(family, m, installed)} {d['about']}; download {d['download_gb']:.0f} GB, uses ~{d['arena_gb']:.0f} GB of RAM{fit}")
- rec = str(names.index(installed or recommended_size(family, names, ram)) + 1)
+ rec = str(names.index(installed or recommended_size(family, names, ram, gpu["vram_gb"])) + 1)
model = a.model or names[int(ask("Which size?", [str(i) for i in range(1, len(names) + 1)], rec, a.yes)) - 1]
budget, q4_split = None, False
in_budget = bool(MODELS[model].get("budget")) and unsloth_budget(model, ram, gpu, a.resident_budget_gib)
@@ -6538,7 +6591,12 @@ def main() -> int:
multi, sel, chosen = [], [gpu["index"]], [gpu]
# (the low-RAM mode's variant is decided once the context is known, below; on several GPUs it is the mapped one
# before RESIDENT_SPLIT_ENGINE, #642)
- if not low_ram and budget is None and ram < MODELS[model]["ram_gb"] - 4:
+ if MODELS[model].get("mem_total"):
+ if mem_have(model, ram, gpu["vram_gb"]) < MODELS[model]["ram_gb"]: # recommend, never force: a warning
+ warn(f"{model} needs {MODELS[model]['ram_gb']} GB of VRAM and RAM together; this PC has "
+ f"{mem_have(model, ram, gpu['vram_gb']):.0f} GB ({ram:.0f} GB RAM + {gpu['vram_gb']:.0f} GB VRAM): it "
+ "may be slow, run out of memory or not start")
+ elif not low_ram and budget is None and ram < MODELS[model]["ram_gb"] - 4:
confirm_paging(model, ram, a.low_ram, a.yes, bool(a.model))
ok(f"size: {model}")
if MODELS[model].get("license"): # the ready-made packs: each size has its own license
@@ -6749,17 +6807,18 @@ def main() -> int:
# ---- 4. the engine
step(4, "the Strata engine")
llama, eng, meta, lib_dirs, gpu, vision = engine_step(a, gpu, hip, amd, cuda_tk, vision)
- if ready and meta.get("source") != "local" and \
+ if (ready or family == "gyro") and meta.get("source") != "local" and \
tuple(int(x) for x in str(meta.get("version", "0")).split(".")[:4] if x.isdigit()) < MODELS[model]["engine"]:
- # a ready-made engine from before NVFP4 cannot read these experts: one compiled from this source can
- warn(f"the ready-made engine {meta.get('version')} has no NVFP4 experts yet: compiling the engine from this source")
+ # a ready-made engine from before NVFP4 / Gyro cannot read these experts: one compiled from this source can
+ warn(f"the ready-made engine {meta.get('version')} has no {'NVFP4' if ready else 'Gyro'} experts yet: compiling "
+ "the engine from this source")
eng = build_engine(gpu, vision, a.yes, llama, toolkit=cuda_tk)
meta = json.loads((eng / "BUILD.json").read_text(encoding="utf-8"))
lib_dirs = meta.get("lib_dirs") or meta.get("cuda_dirs") or cuda_lib_dirs(cuda_tk)
engine_ver = tuple(int(x) for x in str(meta.get("version", "0")).split(".")[:4] if x.isdigit())
need_engine = MODELS[model].get("engine", UNSLOTH_ENGINE)
# checked before the 94-188 GB download
- if MODELS[model].get("budget") and engine_ver < need_engine:
+ if (MODELS[model].get("budget") or family == "gyro") and engine_ver < need_engine:
fail(f"{model} needs engine {'.'.join(map(str, need_engine))} or newer; this one is {meta.get('version')}",
"update Strata (or compile the engine with --build) and run setup again")
ok(f"engine: {eng / EXE}")
diff --git a/strata_control/models.py b/strata_control/models.py
index 30ef2d031..eec730d6f 100644
--- a/strata_control/models.py
+++ b/strata_control/models.py
@@ -33,7 +33,7 @@ MAX_IMPORT_SHARDS = 128
SHARD = re.compile(r"(.+)-(\d{5})-of-(\d{5})\.gguf\Z", re.I)
SUPPORTED_SELECTIONS = frozenset({"qwen/Q2_0", "qwen/IQ2_XS", "qwen/IQ3_XXS", "qwen/IQ3_S",
"swift/IQ2_XS", "swift/IQ3_XXS", "coder/IQ1_M",
- "unsloth/UD-IQ4_XS", "unsloth/UD-Q4_K_XL"})
+ "unsloth/UD-IQ4_XS", "unsloth/UD-Q4_K_XL", "gyro/Gyro-S", "gyro/Gyro-M"})
def settings_path() -> Path:
diff --git a/tools/strata_mcp.py b/tools/strata_mcp.py
index 728958931..dc37773c9 100644
--- a/tools/strata_mcp.py
+++ b/tools/strata_mcp.py
@@ -74,6 +74,14 @@ FALLBACK_MODELS = {
"Q8_0": {"about": "8-bit, the reference quality (every expert Q8_0); for a 192 GB PC",
"download_gb": 188.2, "ram_gb": 136, "arena_gb": 128.3, "families": ("unsloth",), "budget": True,
"vision": True},
+ "Gyro-S": {"about": "the smallest option, for PCs with little memory (27.6 GB of VRAM + RAM together); lower quality "
+ "than the 2-3-bit sizes, and the 2-bit Q2_0 decodes faster",
+ "download_gb": 58.5, "ram_gb": 27.6, "arena_gb": 25.8, "families": ("gyro",), "mem_total": True,
+ "vision": False},
+ "Gyro-M": {"about": "about IQ3_XXS quality with about 20% less GPU memory (35 GB of VRAM + RAM together); decodes "
+ "about 20% slower than Q2_0 on 12 GB cards",
+ "download_gb": 92.0, "ram_gb": 35.0, "arena_gb": 32.1, "families": ("gyro",), "mem_total": True,
+ "vision": False},
"NVFP4": {"about": "4-bit NVFP4 experts, GPTQ (sergqwer's strata-nvfp4 pack); RTX 50 has FP4 tensor cores, other "
"NVIDIA cards run it at 8-bit precision", "download_gb": 128.0, "ram_gb": 80, "arena_gb": 68.0,
"families": ("nvfp4",), "vision": True},
@@ -92,6 +100,10 @@ FALLBACK_FAMILIES = {
"~80 GB (UD-IQ4_XS) to ~136 GB (Q8_0) of RAM, else "
"part of them read from the SSD",
"tag": "unsloth-", "vision": True},
+ "gyro": {"title": "Qwen3.8-Flash-Next (Gyro)", "about": "Gyro-S (27.6 GB of memory) and Gyro-M (35 GB): the smallest "
+ "sizes, for PCs with little memory; 58 and 92 GB "
+ "downloads (the files include the n-gram table)",
+ "tag": "gyro-"},
"nvfp4": {"title": "Qwen3.8-Flash-Next (NVFP4)", "about": "4-bit NVFP4 experts, NVIDIA cards; a 128 GB download, "
"~80 GB of RAM to keep every expert in RAM",
"tag": "nvfp4-"},
@@ -695,6 +707,13 @@ class Strata:
elif low_fits("IQ1_M"):
fam, model, why = "coder", "IQ1_M", (f"{ram:.0f} GB of RAM is below the Coder's 32 GB, but the GPU's "
f"{vram:.0f} GB make up for it (setup's low-RAM mode: slower)")
+ elif "Gyro-S" in models and ram + vram >= models["Gyro-S"]["ram_gb"]:
+ # Gyro (agentionai): only where the sizes above do not fit; the smaller memory total, the 27.6 GB of Gyro-S,
+ # is VRAM + RAM together. Gyro-M (35 GB) when this PC has that much, else Gyro-S.
+ model = "Gyro-M" if ram + vram >= models["Gyro-M"]["ram_gb"] else "Gyro-S"
+ fam, why = "gyro", (f"{ram:.0f} GB of RAM is below the Coder's 32 GB, but RAM + VRAM = {ram + vram:.0f} GB "
+ f"holds {model} ({models[model]['ram_gb']} GB of memory in total; the quality is lower "
+ "than the 2-3-bit sizes, and it needs the whole download on a fast SSD)")
else:
return {"family": None, "model": None, "why": f"{ram:.0f} GB of RAM: Strata needs 32 GB or more "
"(the smallest model, the Coder, keeps ~23 GB of experts in "
@@ -728,7 +747,8 @@ class Strata:
tag = (fd.get("tag", "") + m).lower()
verdict = None
if ram is not None:
- verdict = "fits" if ram >= d["ram_gb"] else "tight" if ram >= d["ram_gb"] - 8 else "does not fit"
+ have_gb = ram + vram if d.get("mem_total") else ram # Gyro: VRAM and RAM together
+ verdict = "fits" if have_gb >= d["ram_gb"] else "tight" if have_gb >= d["ram_gb"] - 8 else "does not fit"
try:
if d.get("budget") and S and hasattr(S, "unsloth_budget") and not S.unsloth_budget(m, ram):
verdict = ("experimental: " if d.get("experimental") else "") + "fits (all of its experts in RAM)"
@@ -738,7 +758,7 @@ class Strata:
if ram >= d["ram_gb"] else
"slow here: part of its experts in RAM, the rest read from the SSD (setup "
"asks first)" if ram >= 48 else "does not fit")
- elif S and S.low_ram_needed(m, ram) and S.low_ram_fits(m, ram, vram):
+ elif not d.get("mem_total") and S and S.low_ram_needed(m, ram) and S.low_ram_fits(m, ram, vram):
verdict = "fits in the low-RAM mode (slower: the GPU holds part of the experts)"
except Exception: # noqa: BLE001
pass
diff --git a/tools/test_setup_menu.py b/tools/test_setup_menu.py
index 59ed91a95..f6778a821 100644
--- a/tools/test_setup_menu.py
+++ b/tools/test_setup_menu.py
@@ -20,7 +20,7 @@ import setup # noqa: E402
class FamilyOrder(unittest.TestCase):
def test_the_menu_order_on_nvidia(self):
fams = setup.menu_families(False) + list(setup.OWN_PATH_FAMILIES)
- self.assertEqual(fams, ["qwen", "swift", "coder", "unsloth", "nvfp4", "qwen36"])
+ self.assertEqual(fams, ["qwen", "swift", "coder", "unsloth", "gyro", "nvfp4", "qwen36"])
def test_the_default_stays_the_first_family(self):
self.assertEqual(setup.menu_families(False)[0], "qwen")
@@ -29,7 +29,7 @@ class FamilyOrder(unittest.TestCase):
def test_nvfp4_is_not_listed_on_amd_and_intel(self):
fams = setup.menu_families(True)
self.assertNotIn("nvfp4", fams)
- self.assertEqual(fams, ["qwen", "swift", "coder", "unsloth"])
+ self.assertEqual(fams, ["qwen", "swift", "coder", "unsloth", "gyro"])
self.assertIn("nvfp4", setup.FAMILIES) # --family nvfp4 still exists, and says why not
def test_the_unsloth_line_names_the_alternative_on_amd_and_intel(self):