MiniMax-M3.1-Flash open weights are now LIVE 🔥🔥 428B params, 23B active. Weights ~252 GB. Link below.

sparkDash 2.70 is out ⚡️ This is a MAJOR release. New pages - Cost compare: see what your fleet's token usage would cost on any hosted model, including cached-prefill savings - Temps: daily and weekly GPU temperature graphs for every unit - Metrics history per unit (48h and 30 days, opt-in), with a History panel on the unit page Engines & hardware - New engines: Strata & Flare (tok/s and benchmarks) and Radiance (live rates) - Intel Arc GPU support - AMD GPU telemetry - GPU clock cap shown on each unit, and you can set it from Edit unit. If the unit has no passwordless sudo, the dashboard asks for the password, or you can run it on the Spark yourself Benchmarks - Decode bench: pick the temperature (default 0.7) and turn thinking on or off for prose. Both show on copied results and the share card - Tool Eval Bench "Show card": a shareable image of a finished result - Comparison card: tick two runs and get one image showing the difference - Compare a run from one Spark with a run from another, plus an expandable scenario list - MAX_ACTIVE_BENCH_JOBS setting to raise the fleet benchmark limit Overview & monitoring - Overview cards list every loaded vLLM model, plus CPU temp, CPU power and the engine's parallel slot count - Token totals for the last 24h, 6h and 1h, with one row per unit and model - Optional card showing inference request activity - Alerts to ntfy, Discord, Slack or a webhook (opt-in), plus a tab badge and desktop notifications Other - "Run both" button in the Models panel - "Back to Overview" link on every unit page - Published container image Link below.

I finally have some numbers behind my DeepSeek V4.1 Flash 1017 quant. EAQ v1 is 11% to 38% closer to the original than LibertAIDAI's EXL3. Seems like the gap is bigger on longer context. To be clear, LibertAIDAI's quant is NOT bad. It's a solid quant, and it gave me a strong baseline to beat. Already in the works: EAQ v1.1, v1.2 and finally EAQ v2. Full details in the model card: huggingface.co/Mia-AiLab/Dee…

The 'old' vLLM recipe for DeepSeek v4.1 Flash *should* work with the new 1017 release. I haven't tried it though. Just try replacing the model. That being said, I'm working on a TensorFold recipe. vLLM recipe here: github.com/MiaAI-Lab/DeepSee…

sparkDash 2.30 is out ⚡️ Windows PCs as units Monitor any Windows PC with an Nvidia GPU over its built-in OpenSSH. That covers RTX Spark machines, which run Windows, and any other Windows box with an Nvidia card. RoCE / RDMA panel - Link and traffic: link state, rate, MTU, RX/TX traffic and interface errors. - PFC and QoS: which priorities have PFC on, the trust mode (PCP or DSCP), the DSCP map when DSCP trust is on, and the global pause settings. - Counters: a collapsible list of RDMA counters, plus ethtool discards, CRC errors and pause frames. It shows the change since the last sample. What else is new - Installable mobile app (PWA) with an in-app install prompt - Permanent monthly energy history, with a month/year view - Timezone settings - Prometheus /metrics export (opt-in) - Per-GPU graphs on multi-GPU hosts - Model picker in the decode benchmark for multi-model servers - Per-Spark llmHost for LLM probes and benchmarks - Restart sparkDash from the UI when the fleet changes - Tooltip on “Partial coverage” naming the nodes with no fresh reading - FreeToken KV cache, Running and TTFT tiles on the LLM panel - README guide for running the dashboard on a non-Spark machine Fixes - SSH tunnel fallback for benchmarks and Showcase - Poll interval setting now really sets the polling rate - "Frontend not built" after a Docker upgrade - Blank page on a token-protected remote bind - Model launchers no longer inherit PORT=5555 or the access token - Two-unit Overview row stretches to the full width - Efficiency on the Overview card keeps its digits - Add/Edit dialog wording follows the unit type - Clearer SSH error when a Windows PC is added as a Linux unit - Dev server listens on loopback only 30 issues and PRs closed. Thanks to every contributor who opened an issue or PR, or shared feedback here on X.

As I said many times before, TensorFold is a real turning point for local AI. Just try it for yourself, we have a recipe for everyone! All of my TensorFold recipes can be found here: mia-ai.net/models?for=engine…

GLM 5.3 Flash on 2x DGX Sparks v1.10 is out ⚡️ Lots of PRs merged and bugs squashed! Here are some highlights: - Works with Claude Code & Anthropic SDKs directly - Much faster at very long chats (500k+ tokens) - Opt-in disk cache: long chats resume in 0.2 s instead of 11 s, even after a restart - Sampling is now as fast as greedy decoding Full changelog here: mia-ai.net/models/GLM-5.3-Fl…

I went through every GLM, Qwen, Kimi and DeepSeek release to predict the next release dates. GLM (5.4 or 5.5): Oct 13-22 Qwen4: Oct 12 - Nov 15, with 27B as the priority Kimi K3.1: mid Oct to mid Nov, then K4 around Jan DeepSeek v4.1 Pro: Oct 20 v Nov 30, with V5 in 2027 China's Golden week ends tomorrow, and Chinese labs usually ship right after it. The next 6 weeks are going to be busy for local AI 🔥

The new GLM 5.3 Flash EXL3 quant is calibrated differently. You don't have to believe me, try it yourself. HF link: huggingface.co/Mia-AiLab/GLM… Recipe: mia-ai.net/models/GLM-5.3-Fl…