Pocket Budgets, Wish-Driven Sprints: DeepSeek Refresh 1017 Verdict

A dated refresh landed on October 17, and it skips retraining entirely. DeepSeek tuned behavior through post-training alone, then pitched the outcome as wish-driven programming: describe a project, receive a working build. A hands-on trial paired refresh 1017 with Hermes , a helper harness, and ran portal clones, sites, plus games through it.

The kitchen story stays simple. DeepSeek V4.1-Flash holds 552 billion mixture parameters, yet each request wakes a sliver: 8 billion while reading input, 16 billion while writing output. Picture a vast library where every question opens two shelves. A million-token window plus native image reading completes the package, and the official DeepSeek announcement spells out the same shape.

How memory bills shrank

Speed plus thrift starts in memory. The model shrinks the bloated scratchpad behind helper workloads to one quarter of fast memory and one eighth of disk. Think smaller warehouse, identical output. Because cache hits dominate helper bills, the cut lands straight in the invoice. An arXiv compression paper shows the arithmetic, down near 890 bytes per token.

Refresh 1017 moved the scoreboard too. Per the DeepSeek changelog, Terminal-Bench 4.0 reached 42.3, DeepSWE hit 77.0, and HLE sans tools hit 42.6; visual exams climbed as well. Rigs included Claude Code scaffolds plus Mini-SWE-V2. Weights sit on HuggingFace as a 48-shard open bundle, and a MindStudio recap mirrors the same table.

The showpiece stayed a portal clone. A client page selling service packages reappeared on a local host, reportedly inside ten minutes, with one message finishing inside five. Striking, yet solitary: the only evidence remains the creator's own stopwatch. Still, the wish-driven claim looks substantive rather than hollow.

Identical brains, divergent hands

Identical brains, divergent hands. The trial set Claude Code beside Hermes using refresh 1017, judging the Hermes render livelier. No shock there: surrounding software shapes outcomes as much as weights do. DeepSeek itself spotlights its own harness alongside CodeBuddy plus WorkBuddy, requesting release 0.2.1-rc.1 or newer. OpenCode support ships on the same list.

Money claims deserve slow reading. Three builds allegedly cost 38 cents where matching tokens on Opus would bill 21 dollars, averaging fourteen minutes. Those figures come from one workbench, unverified elsewhere. Official framing helps: DeepSeek prices quiet hours at half rate; OpenRouter lists live unit prices, while a Vals AI page frames cost comparisons. Thrift looks real, but fifty-fold ratios need home measurement.

Visual chores split the verdict. A 3JS diving game, four sites from one command, plus a 3D racer all emerged; official wish-driven samples add a Forbidden City model alongside a racing game. Yet the creator draws an honest boundary: image polish trails Opus 5.5. Fast plus thrifty, then, just not chart-topping on showcase polish.

One absence nags: rival GPT-6 Astra never enters the official table, a gap the trial flags too. Science-heavy helper chores still favor giant systems. The thesis stands narrow: nimble plus swift covers perhaps 95 percent of daily chores, giants keep the residue. That percentage voices a vendor claim, never independent lab proof.

Who should try? Small squads, students, tinkerers: grab the HuggingFace bundle, open a 1017 profile inside OpenCode or the DeepSeek harness, and treat early runs as free rehearsal. Search plus Fetch rides free through helpers until October 31, metered afterward. Measure two weeks on home workload before trusting production secrets. Revisit live docs before deciding.