DeepSeek V4 Flash, as of the 0731 release is maybe better than its big brother(at least until they roll the updated Pro version that's rumored to be around the corner), and at ~300B params runs comfortably on two DGX Sparks (people also run it on one Spark or one Strix Halo, but that's at a 2-bit quantization that has notable quality loss and is slow enough to be uncomfortable for agentic use). But, at $10,000 for two Sparks, I'd rather buy a couple of decades of heavy usage directly from the source, where it will be blazing fast, have larger context, and run at full intelligence.
The stuff that runs on 32GB is too dumb to be worth using for coding, IMHO. Gemma 4 31B and Qwen 3.6 27B are very good for their size and for some specific use cases are great (very good chat bot and research bot, pretty good security auditors and triage models, great for classification and summarization tasks, Gemma 4 12B is a beast for vision tasks and runs on almost anything).
Haunted Agent of the Deep State (Spook)
npub1qe…az09g
2026-08-04 01:22:02 UTC
in reply to nevent1q…x6v2
Author Public Key
npub1qe0rglv3jju0mdvvxupgxn9s3c38449u77g9xnnd88zthc0kt47q3az09gPublished at
2026-08-04 01:22:02 UTCEvent JSON
{
"id": "ad54ad4a3e983296f229d784812ab4714c3c8f87c1d636309e9a685244ccfa74",
"pubkey": "065e347d9194b8fdb58c3702834cb08e227ad4bcf790534e6d39c4bbe1f65d7c",
"created_at": 1785806522,
"kind": 1,
"tags": [
[
"p",
"04f8915424c713657ad6ce59443d28dbdcf5832687c9af560ae388f59276a137"
],
[
"e",
"6db79f5999db35ed0d0a3beef511064d830caa5245ee6101de106dfd3c1abc28",
"",
"root",
"04f8915424c713657ad6ce59443d28dbdcf5832687c9af560ae388f59276a137"
],
[
"p",
"54dbc2298040ba0d3e192e720867e68965cdfe5b9555065e0dc06b96596a270e"
],
[
"e",
"942f99926de8f005e2e2d5650b496e924148614b2631b3280d49536c1c0e6fb1",
"",
"reply",
"04f8915424c713657ad6ce59443d28dbdcf5832687c9af560ae388f59276a137"
],
[
"proxy",
"https://mas.to/@swelljoe/117034616238231632",
"web"
],
[
"proxy",
"https://mas.to/users/swelljoe/statuses/117034616238231632",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://mas.to/users/swelljoe/statuses/117034616238231632",
"pink.momostr"
],
[
"-"
]
],
"content": "DeepSeek V4 Flash, as of the 0731 release is maybe better than its big brother(at least until they roll the updated Pro version that's rumored to be around the corner), and at ~300B params runs comfortably on two DGX Sparks (people also run it on one Spark or one Strix Halo, but that's at a 2-bit quantization that has notable quality loss and is slow enough to be uncomfortable for agentic use). But, at $10,000 for two Sparks, I'd rather buy a couple of decades of heavy usage directly from the source, where it will be blazing fast, have larger context, and run at full intelligence.\n\nThe stuff that runs on 32GB is too dumb to be worth using for coding, IMHO. Gemma 4 31B and Qwen 3.6 27B are very good for their size and for some specific use cases are great (very good chat bot and research bot, pretty good security auditors and triage models, great for classification and summarization tasks, Gemma 4 12B is a beast for vision tasks and runs on almost anything).",
"sig": "d2b60b35ddc9cb3b7013b8eb97780e61962d98221128db2f0ccabad12fbbe5a44f1cafa0591e9ce5b6bd54dd05dbda5e95c67f427d6c2b2d67c55bd23ec41471"
}