Seven Qwen3.8-27B Recipes, No Single Winner on a Spark

@melissapan and collaborators evaluated 7 models on Claude Code, Codex, and Pi. On SWE-bench Lite, Fable 5 hit 97.8% accuracy at $1.33 per rollout in Claude Code against 96.7% at $0.67 in Pi, twice the cost for a 1.1-point gain. Pi reached the Pareto frontier with four tools: read, write, edit, and bash. Claude Code’s mean initial context ran over 10x Pi’s, and GPT-5.6 Sol scored higher in Pi than in Codex.

@JacquelineSYC19 describes how Artie’s 17-person team shares Hermes profiles for engineering, sales, design, and BizOps, with skills versioned in Git and a human reviewing every PR before it merges. The box cost $60 a year at Hetzner. Each night the agents “dream”: a cron job reviews the day’s work, promotes useful lessons into memory, and emails the engineer a report. Evals run on self-hosted Arize Phoenix with Datadog underneath.

@burkov tested DeepSeek 4.1 Flash in OpenCode and calls it a good coder that “doesn’t easily see the bigger picture”: every requirement must be stated before it will code, and role-play prompts such as “imagine you are a X” get taken literally. @buswe_com replied that writing the plan to a file with steps and file paths worked better than a role.

Google released Gemma 4 in four sizes, E2B, E4B, 26B-A4B, and 31B, all Apache 2.0 and all multimodal. @Sumanth_077 reports AIME 2026 math at 89.2%, up from 20.8%, and LiveCodeBench at 80.0%, up from 29.1%. His Unsloth guide trains E2B LoRA on 8-10GB of VRAM, E4B LoRA on 17GB, and the 31B with QLoRA on 22GB, 1.5x faster with 60% less VRAM than standard training. Gemma 4 supports 140 languages.

@Blackwellboy reports that Victor, posting as @ViC305, runs DeepSeek-V4.1-Flash on a single 128GB Spark at up to 24.7 tok/s with 1.59-bit SAGE-EXL3 experts, 91.2% top-1 agreement with the native FP4, and 97.9% on code. Model load takes about 40 seconds. @jvr0x adds that NVIDIA published an official NVFP4 version of DeepSeek V4.1 Flash on Hugging Face at 491GB. One commenter, @jayleaton, could not get it running outside a headless setup.

@wei_wang reports the dual DGX Spark DeepSeek V4.1 Flash setup sped up again: single-path prose went from about 32 tok/s to 42, and two-stream throughput from 42.5 to 63.7 tok/s, gains of about 33% and 50%. The new Cooperative MoE kernel batches the small expert matrix operations that MoE decoding spawns at low concurrency. Weights are unchanged, still EXL3 at 2.9 bpw, and 32K prefill barely moved. Real code tasks gained about 11% and inference about 29%.

@anderslie argues Jev’s speed comes from the inference path: label the choices, prefill the shared context, fork it per question, and read logits constrained to valid labels in one parallel pass. Jev’s own model, he notes, is trained to perform well in exactly that setting. @harshagundal shipped an open version, Qwen-2.5-1B-RLCD on Hugging Face, claiming 5x faster on-device inference for type-safe JSON workloads with no new training, demonstrated on an M4 MacBook and later playing Doom at 150ms latency.

@redp314 wired Jev into pull-request review: one call to @typesafeai per diff returns 14 typed checks as probabilities, from hardcoded secrets to broken APIs, and anything between 0.35 and 0.65 confidence escalates to a human or a bigger model. He reports 6 real PRs at $0.00007 each, 1,000 PRs for 7 cents against roughly $14.50 on Opus 5, with answers in half a second.

@jun_song frames the release as the pace story: while the US argues about pacing the frontier, companies across China keep shipping open weights, following Meituan. Citing @ModelScope2022, he reports China Telecom’s Xing4.0-29B-A4B is a 29B MoE that activates 4B parameters for long-horizon agents, licensed Apache 2.0, and leads both comparison models on Claw-Eval and Terminal-Bench 2.1.