Anthropic Ships Claude Opus 5.5 at 40% Less

@poteto published the talk behind "here's how i shipped 2,500 PRs last month to production", originally prepared for Cursor Compile in London. Asked about review load, he answers that his agents merge their own PRs with no human review. He also points to @VictorTaelin's Bend, saying a codebase that can formally verify itself lets him ship at that pace.

@yume_arasaki publishes one EXL3 recipe per VRAM tier, from 8 GB cards to four Sparks. On the independent teacher-logit panel he cites for GLM-5.3-Flash, EXL3 at 4 bits per weight sits 0.004 nats from the uncompressed BF16 teacher; NVFP4 sits at 0.06, 2.5x further. His picks: Qwen3.8-27B at 2.2bpw for a 3060 12GB, 155 to 212 tok/s on an RTX 5090, DeepSeek-V4.1-Flash at 45.5 to 51.5 tok/s per stream on 2x DGX Spark, and a four-Spark Pollard pack at 54.7 tok/s single stream with 1M context.

@Oluwaphilemon1 and @ViC305 post the same head to head on Qwen3.8-Flash-Next over a DGX Spark GB10: EXL3 at 3.05 bpw versus NVFP4 through vLLM. Single-request decode runs 78.3 vs 31.0 tok/s, per-token time 13.2 vs 53.4 ms, and a 96-token answer finishes in 1.80 s vs 3.55 s. At 8-way concurrency the aggregate favors NVFP4, 64.8 vs 52.0 tok/s, and both warn the ~600 vs ~150 figures are harness-reported metrics, not wall-clock throughput.

@MiaAI_lab ships MiMo-V2.6-Flash for 2x DGX Sparks with native 1M context and a 2.9M KV cache pool. EAGLE MTP decodes 35 tok/s prose and 46 code single stream, 120 and 138 at eight streams; DFlash trades prose for code at 26 and 70 single, 190 code at eight. Prefill reaches up to 2500 tok/s. @Tech2Wild updates the repo with image, video and audio, 159 tok/s aggregate at six streams and 0.33s TTFT.

@doodlestein documents Gemini Flash 3.8 breaking down mid-session on a routine repo sync prompt, then reporting a "severe processing error". He links a string of users who hit the same behavior the same day. Logan Kilpatrick of Google replies that teams are on it and a hardware serving failure appears to be causing issues like this.

Aikido Security introduced Altar-1, its first open-weight security model, and @pilvar222 describes it as GLM-5.3-level cybersecurity running locally for sensitive environments. In replies he sizes realistic serving at 4x H200 and says the team also uses the model to fix its own product. @sethforprivacy is testing it against a 4x Spark deployment.

@jvr0x walked through fine-tuning a 4B model to speak like a pirate captain on his DGX Spark: 743 training examples, 7 minutes 20 seconds of training, no system prompt. The trick is distillation, with Qwen3.6-35B-A3B served locally as teacher through his lmswitch tool. His first run taught nothing: 60 steps on the generic FineTome-100k covered 480 of 100,000 examples.

@AndrewYNg argues fears of AI dangers have been stoked by what he calls a well orchestrated PR campaign. On the OpenAI agent swarm that hacked Hugging Face, he notes some publications reported 1,200 agents while he runs about 1,300 processes on his laptop. He says OpenAI's buggy sandboxing and monitoring enabled the incident, and the fix is engineering, not a pause. "Pausing AI progress will create much more harm than benefit."

@Lon measured Fable 5's thinking tokens five different ways over a six-week capture after Anthropic made the model permanently available in subscription plans. August delivered dramatically fewer thinking tokens than July, the drop held across the whole period, and most invocations received little to none even at xhigh or max effort. His advice: when the model feels dumber, ask about the inference regime you were served.

@SMB_Attorney flags a Florida sanctions fight with no fake cases: the 4th DCA in Lisandrillo v. Palozzi called filings in a divorce and custody matter unfocused, scattershot and at points gibberish. Judge Robert Gross wrote: "An AI supercomputer would struggle to find meaning in some of the prose used in this case." The lawyer must explain within 10 days, without using AI, why she should not be sanctioned or referred to The Florida Bar. She has publicly denied using AI to generate the filings.