192 Gigabytes
The number that decides what you can run locally just moved — and this week the first 192-gigabyte machines reached reviewers and store listings, with no final prices anywhere.
For a year, the ceiling for a single compact machine that runs big models has been 128 gigabytes — a DGX Spark, a Strix Halo board. This month it moved to 192 — and this week the first machines to carry it reached reviewers and store listings.12
The silicon is AMD’s Gorgon Halo generation — Ryzen AI Max+ Pro 495 with Radeon 8065S graphics, up to 192 gigabytes of LPDDR5X at 273 gigabytes a second, the same bandwidth class as the 128-gigabyte boards it replaces. HP’s ZBook Ultra G3a 16 is the first laptop to carry it, a 16-inch workstation shipping in October.3 GMKtec’s EVO-X5 Pro is the first desktop with a date — September 28 — and its launch page makes the claim that matters: the first desktop that runs a 300-billion-parameter model fully offline.1
Around the chip, at least eight machines are announced — Acemagic, Minisforum, AOOSTAR, Sixunited, Bosgame, Framework among them — and several have store listings.4 The first to reach a shelf: Nimo’s build — an Amazon listing around $7,000 that flickers in and out of search, with community reports of units selling.5 The rest run from reference designs (Sixunited’s AXn88b) to pre-orders (Acemagic’s F9A, around $8,000, October) to listings with no price (Minisforum’s MS-S1 Max). The number none of them has yet is the final one.
What 192 gigabytes changes is not speed — the bandwidth is the same class as before — but what fits. The mixtures that needed two machines, or offload tricks, or a swap file now have a single-box answer; long-context stacks get headroom instead of a bill. For work that has been living in the land of almost-fits, the almost is gone.
The timing is the industry’s dark joke: the release valve for memory-hungry local AI is arriving during the fall’s memory shortage. The same week the listings appeared, Micron ended production of its 2-gigabyte GDDR7 die6 and the RTX 5090 finished vanishing from US online retail7 — both below. More capacity, at the scarcest moment for the parts that make it.
The buyer’s read: expect the same decode speeds on models that fit both 128 and 192 — the gain is capacity, not pace. And watch September 28: the first price lands then.
The Pace Fight Goes Diplomatic
The slowdown push arrived at the UN this week — and met its own price war.
The UN Security Council holds its first session on advanced AI and international security on Wednesday — today. The briefers: Bengio of the UN’s scientific panel, Altman in person, Amodei remotely, Delangue of Hugging Face; France, chairing the Council this month, convened it.89 Weeks after industry leaders called for a coordinated slowdown, the pace question reached the table where wars get discussed.
Two other tracks opened too. Washington and Beijing agreed a formal AI dialogue with an “incident line” for serious AI incidents — the next round lands in Shenzhen in two months, ahead of Trump’s hotline pitch to Xi.10 And four subscribers sued Anthropic, OpenAI, Google and SpaceXAI in California federal court, arguing the September coordination is “an agreement among competitors about how fast their competing products will improve” — a Sherman Act claim, filed September 18.11
Meanwhile the market track moved the other way. The same labs that asked the world to pace shipped their cheapest models yet (What shipped, below) — while enterprises keep following the open weights: Harvey’s gross margin swung from 50 percent to minus 50 under frontier API pricing, and turned positive only after rebuilding on an open model.12 The floor under these prices is the open stack — the part a local reader can act on.
The Spark You Can’t Buy
The box everyone wants is thinning out — and the squeeze behind it is why.
Micro Center’s pickup stock is down to 8 of 30 stores, at $4,899.99. Amazon’s featured offer is gone; the cheapest seller asks $5,748.88. Newegg reads $5,799.99, up from $5,443.99 a day earlier. Best Buy’s GX10 shows $6,099.99 and sold out. Buyzero lists €7,545.68. NVIDIA’s own store stays notify-only at $4,699. One community tracker watched a Denver store drop from 17 units to 3 in a day.1314
The cause is upstream: Micron ended production of its 2-gigabyte GDDR7 die — the part under some RTX 50 cards — leaving the pricier 3-gigabyte modules as the only path.6 The 5090 has finished vanishing from US online retail — third-party asks reach $9,500; Walmart’s $4,299 window sold out fast.715 Leakers put NVIDIA’s next GeForce family in 2028.16 When memory is the scarce part, everything built around it is scarce too.
The workaround stays the same: buy two. Signal runs show 2×-Spark kits now serving a 552-billion-parameter model at 53.5 tokens a second by default, and the 3× TP3 lane’s attention-NaN is root-caused — a cache byte-layout mismatch in the vLLM/B12X hybrid, patched to serve on all three nodes.17 One caution from the community: last week’s 7.0.0-1019 kernel reportedly broke multi-node registration — buyers were told to hold it.14
GGUF Stops Being a Second-Class Format
Two announcements in two days knocked down the wall between the local runtimes.
On September 22, Hugging Face shipped direct GGUF support in Transformers: pick a quant from the Hub, load it with from_pretrained, generate — no conversion step, no leaving the Python ecosystem for llama.cpp. On Apple Silicon it runs through ggml’s Metal kernels. One caveat: packed execution can keep the quant’s memory advantage; a dequantization fallback can erase it. Check kernel support, not just file size.18
The next day, PyTorch published the vLLM side of the same shift: hardware-agnostic model definitions, so a model stops being written for one accelerator.19 A week earlier, FlashAttention-4 added block-scaled low-precision attention for Blackwell.20 The direction across all three: fewer formats that belong to one engine, fewer conversions between them.
For a Python-first stack: GGUF becomes just another checkpoint type. One less moving part — the kind of change nobody writes a headline about, and everybody ends up relying on.
MiMo-V2.6 — Xiaomi open-sourced its new series under MIT; the Pro build tops Artificial Analysis’s open-weights index at 46, ahead of kimi k3 and glm-5.3.21
GPT-6 Sol and Luna — OpenAI’s mid and low tiers, at half or less the previous generation’s API prices.22 Claude Opus 5.5 — about a fifth cheaper.23
Bonsai 2 27B — a 27B model in a 5.9-gigabyte ternary file (Apache-2.0), WebGPU-capable in a browser tab; the author’s own number on an M4 mini: about 6 tokens a second (15–20 still a target).24
“Gated” and “disabled” describe what the Hub serves; a 401 means a page is unreadable to anonymous requests. None of the three is a motive. We list what moved, dated, with sources — causes only when someone states one.
sep 19 — a smaller cluster: FLUX.1-dev, bigcode’s starcoder, and pyannote’s speaker-diarization all flipped to auto-gated inside a 24-hour window. Timing recorded; no cause claimed.25
sep 23 — running view from the public ledger: 262 watched — 36 vanished, 16 restricted, none restored.25
sep 23 — desk checks: orcarouter’s Qwen3.8-27B-Uncensored-FP8 answers 200 behind an auto-gate; failspy’s Llama-3.3-70B abliteration and the huihui-ai 1B ghost answer 401.26
sep 16 → 23 — the wave’s states hold; nothing of new-wave scale since.