New GLM 5.3 Flash Quant Arrives for Two DGX Sparks

@MiaAI_lab's 4-bit quant cuts 4-bit expert error by 28-46%; two-Spark builds decode at 57.4 tok/s.

@MiaAI_lab shipped a new 4-bit EXL3 quant of GLM 5.3 Flash built for TensorFold, the same format and size as the TR3 4bpw build most owners ran. @plotarmordev reports the calibration removes 28-46% of the error the 4-bit experts add and sits 2-18% closer to the full model on 7 of 8 test sets. The lab counts up to 37% fewer confident wrong picks.

@StevenDarlow put the weights on two DGX Sparks and recorded 57.4 tok/s median decode standard and 55.3 abliterated. An earlier run held 53 tok/s across a 1M-token context on the abliterated build. The recipe, the Hugging Face weights and PR #52 are public, so an owner with two Sparks can reproduce the numbers.

The lab says the quant is equal in coding, better in everything else, and thinks about 10% less than the old TR3 build; replies on coding tests run about 10% shorter. @MiaAI_lab answered a calibration challenge by posting that every full-precision tensor is byte-identical to zai-org's BF16 checkpoint, verified across all 1,618 of them.

From posts by @MiaAI_lab, @plotarmordev, @StevenDarlow

Models & Hardware

Full GLM 5.3 Runs on Three DGX Sparks

@MiaAI_lab reports 24-28 tok/s on prose and says a four-Spark build is next.

The full GLM 5.3 model, not the Flash variant, now serves from three DGX Sparks under TensorFold at 24-28 tok/s on prose, @MiaAI_lab reports. Asked about a two-Spark tensor-parallel build, the lab answered “Won't happen”, while a four-Spark version is promised. @alexellisuk noted an earlier run topped out at a 32k maximum context and a 44k pool.

From @MiaAI_lab

Industry

A Law Firm Open-Sources Its Templates in LLM Markdown

General Legal posts attorney-drafted contracts as CC0 markdown built for models, alongside the original .docx files.

The attorneys at General Legal released 12 template sets under CC0 1.0, from advisor agreements to privacy policies. Each ships as LLM-optimized markdown that flags the fields to customize, plus the original .docx files. JP Mohler, @jaypmo, says the firm's lawyers and deal flow are the IP and calls the precedent documents “underlying scaffolding to enable the work”.

From posts by github.com, @jaypmo

Tesla Plans a 10 Million Robot a Year Factory

@TheHumanoidHub maps the 7 million sq ft Giga Texas site, with initial Optimus production planned for 2027.

@TheHumanoidHub reports the Optimus factory at Giga Texas will cover 7 million sq ft with a targeted capacity of 10 million units a year, and initial production is planned for 2027. The Fremont line, planned at 1 million a year, is expected to begin production this year. @remountco notes the target works out to about 19 robots a minute.

From @TheHumanoidHub

A 13-Step Playbook for Vibe Manufacturing

@gregisenberg maps how frontier models and fabrication services turn a design into a shipped product.

@gregisenberg's 13 steps run from digital calipers and a Polycam scan to a frontier model like Claude Opus 5.5 or Astra writing the design in OpenSCAD or CadQuery. He prices a home printer at $300 to $600, sends files to SendCutSend, Xometry or JLCPCB, and pays an engineer for one review hour. @nikitabier's quoted example: a wifi-controlled dog door, vibe-fabricated to his house specs.

From @gregisenberg

Report Says Anthropic Co-Founder Nearly Walked From Vatican Event

Christopher Hale cites a New York Times report on Chris Olah, Pope Leo XIV, and machine consciousness.

@ConceptualJames passes along Christopher Hale's post citing a New York Times report. It says Anthropic co-founder Chris Olah threatened to walk out of Pope Leo XIV's AI encyclical launch in May, because the pope rejected the idea that machines can be conscious. Olah's team then lobbied privately. Replies ranged from @TrackingFires mocking the walkout threat to @afdevine_ calling it a move to shift liability.

From @ConceptualJames

Sources

Friday, October 2, 2026 Back to the Archive Sunday, October 4, 2026