Qwen3.8-Max-0902 Debuts Atop Code Arena With 1M Context

@mardehaym says 98% of the code at Limestone Digital isn't handwritten. His factory gates every ticket behind an approved spec and plan; review is a five-step pipeline capped at five comments per PR, each a verbatim snippet of five lines or fewer. On a logistics client's rebuild, cycle time dropped 50% and 122 PRs merged in 90 days, 104 with fewer than five comments. A companion post ships three systems for a B2B lender, ACH funds movement, underwriting and a syndicator portal, with 2 engineers and zero disruption.

@kartiksmath argues a harness is just a loop around a stateless LLM, that the right system prompt and tools reproduce any harness, and that prompt caching, which bills uncached input at 10x on Claude, pushes every harness into the same append-only loop. Terminal-Bench results back the toss-up: GPT-5.5 scored 84.7% with NexAU-AHE and 83.1% with Capy against 82.2% with Codex on Terminal-Bench 2.0, while Terminal-Bench 3.0 is topped by Opus 5 in mini-SWE-agent at 42.7%. Compaction, he writes, is the one place harnesses still differ.

@liam_fallen built a business dashboard in Cursor that lives locally on his Mac, then gave a Grok Bot a standing job: pull revenue from Stripe, traffic from Analytics and Search Console, and spend from Meta and Google Ads, clean the mismatched formats, and update the page each morning. Anything odd goes in a section called 'Things worth looking at'. He watches the first runs, checks the bot's pulls against the original dashboards, and only then turns it into a routine.

@MichaelGannotti ran DeepSeek V4 Flash Vision-Exp against GLM-5.3-Flash-EXL3 R4 on the same pair of DGX Sparks on 2026-09-01, same 5-category plus vision harness, same formal OpenAI tools path. DeepSeek scored 25/25 to GLM's 24/25, decoded 78.7 tok/s structured and 28.5 tok/s prose against 53.0 and 18.4, and booted a 1,048,576-token context where GLM managed 540,000. GLM R4's weights came off the cluster to reclaim disk, and DeepSeek Vision-Exp is the live pin.

@MiaAI_lab updated DeepSeek v4 Flash Vision for DGX Sparks with serving fixes: TTFT spread across 4 streams down 38%, aggregate throughput of 4 streams up 12%, per-stream decode spread down 17%, and worst-stream TTFT 2.8 seconds lower. A new TP=3 path carries about 5M KV cache; decode is faster on TP=3 while prefill runs a bit slower, and she says it feels snappier overall.

@Blackwellboy tested Victor Cruz's ~95 GiB EXL3 MixedK DeepSeek V4 builds on a single DGX Spark against his 2-Spark Keys setup, covering quality, tools and speed. @ViC305, who built them, reports the verdict: intelligence and long generations still go to 2 boxes, but 1-Spark MixedK is 'shockingly competitive' for local agents and leaves a whole Spark free. He also notes the test ran with MTP broken and hints a patch is coming.

@redp314 published the full recipe for about 70 tok/s Qwen3.8-27B on one DGX Spark, a speed he says holds from 320 tokens to 65k of context and calls the fastest published for this model on a single Spark. The repo ships an SGLang launch script with every knob explained, a Dockerfile, a 4-bit drafter setup, six benchmarks with CSVs for every number, and a kernel-tuning cache draw worth +16%. The README carries a prompt so Claude Code or Codex can download, build, launch and verify it unattended.

@stevibe spent $4,699 on a DGX Spark, used it as a chatbot for six months, then built SparklingKit after realizing 128 GB of unified memory holds six small specialists at once instead of one generalist. The open-source panel chains transcription, OCR, translation, image grounding, text-to-image, mind maps and chat, queues a 3-hour recording, and keeps everything as plain files in one folder with no lock-in. He argues a 1.7B speech model transcribes better than a 35B chat model ever will.

@helmcode priced a GLM-5.2 FP8 deployment on 8x H200: about €21,500 a month to rent the node at $4.30 per GPU-hour, plus €3,750 for 0.5 FTE of MLOps, €25,250 all in. Break-even arrives at about 14 billion tokens a month against standard API prices, about 25 billion against competitive ones, and never against the current promotional price, which undercuts even a full node. Buying the hardware instead means roughly $300,000 to $420,000 up front for the node, plus power, cooling and people.

@cb_doge posted a Grok Bot summary of Elon Musk's G20 address. It cites a consensus of at least a 15 gigawatt shortfall of power for AI chips in 2027, AI chip production rising about 40-50% a year against power outside China rising 10-20%, and Google and Anthropic already leasing compute from SpaceX because it built its own power plants. The same summary puts digital AI's lift to the global economy at 20-30%, about $20-30 trillion a year.