What happened
In pull request #477, Project Maya collaborator needmorevram proposed a third model family for Maya: MiniMax-M3.1-Flash, served by a new engine called maya-mm3. The pull request describes support for a community EXL3 quantization, Mia-AiLab/MiniMax-M3.1-Flash-EXL3-EAQ-v1-3.5-bpw, listed there at 196.7 GB. Maya’s setup script offers the model beside GLM and DeepSeek. The PR says the change rebased onto Maya v1.1.27; the supplied source excerpt also reports that it was merged and shipped in v1.1.29.
The engine is designed around the model’s sparse-attention and mixture-of-experts structure. It reads EXL3 weights in place, profiles experts so frequently used ones can be held in GPU memory, and keeps other experts in pinned system RAM for CPU work or transfer. Tensor parallelism is configurable with --tp 1|2|4; the implementation groups GPUs, distributes parts of the model across them, and coordinates work through the host. The engine also adds DSpark speculative drafting, an integrated vision tower for pictures, and the model’s five thinking-effort levels. The PR says video input receives a clear error rather than being handled as an image.
Two optional KV-cache features address memory pressure at long context. INT8 KV cache reduces storage but is not exact: on 16K tokens of assistant text, the author reports perplexity rising from 6.996 to 7.009, so BF16 remains the default. KV streaming keeps a configurable recent window in VRAM and stores positions in mapped pinned RAM; the author reports byte-identical logits in a 6K comparison and matching perplexity across 7,972 positions in an 8K test. The setup enables streaming and selects tensor parallelism when the available GPUs form supported groups. The dashboard exposes expert placement, GPU allocation and KV status.
The reported performance comes from one Linux PC with an RTX 3090 and eight RTX 5060 Ti cards, 152 GB total VRAM, and 183 GB RAM. At 8K context, --tp 4 with DSpark reached 38.2 tokens per second on code and 32.4 on prose; prompt processing was about 99 tokens per second in a 6K test. At 128K context with INT8 KV and streaming, prompt speed was 84.1 tokens per second on a 53K prompt. The PR’s checks include reference comparisons, long-context needle retrieval and repeated tensor-parallel runs. Its accompanying excerpt says the author tested Linux on sm_86 and sm_120 hardware, not Windows or AMD.
What the archive adds
This is a new fact on a known Project Maya story: Project Maya is an MIT-licensed inference stack that runs the 321-billion-parameter GLM-5.3-Flash mixture-of-experts model on Linux workstations described the stack’s GLM deployment. The new pull request extends that local inference setup with a separate engine for MiniMax-M3.1-Flash; it does not replace the existing GLM or DeepSeek engines. Earlier coverage also noted MiniMax M3.1 Flash’s announced API and other availability. This is specifically about a community quantization and local runtime support.
Trend signal
This supports harness architecture outlives model choice: Maya is adding a model-specific engine, setup, serving and dashboard path while retaining other model engines. The evidence is bounded to one open-source inference stack and a pull request; it does not establish that this architecture is dominant across production deployments.
Why now
The pull request was opened on October 25, 2026 and closed on October 26, according to the supplied source material, which says it shipped in Maya v1.1.29. That timing makes this a concrete availability update for Maya users. The source does not explain a broader market or technical event that prompted the release.
Why it matters
For teams that already run Maya on Linux hardware, the change makes MiniMax-M3.1-Flash a local option within an existing setup rather than requiring a hosted endpoint. The implementation’s combination of GPU placement, CPU expert execution and streamed KV storage illustrates the engineering trade-offs involved in fitting a large model onto a workstation-class system. Its measurements are useful as a reference for that specific machine and configuration, not as a basis for estimating performance on different hardware.
The work also makes the serving layer part of the model choice. Supporting a new model required more than downloading weights: Maya added its own engine, tokenizer handling, chat template integration, server behavior and dashboard reporting. Builders considering local deployment should account for that ongoing model-specific engineering and test quality, compatibility and speed in their own environment before relying on a new backend.
Limits and uncertainty
The supplied source is a pull request and its implementation notes, not an independent evaluation. Performance numbers come from one Linux PC with a particular mix of GPUs, memory and PCIe connections; they cannot establish typical speed, power use, cost or performance on other systems. The quality checks described are implementation-specific comparisons and tests, not a broad assessment of model capability. INT8 KV is explicitly approximate, while the source’s exactness checks for streaming cover stated test cases rather than every workload. The supplied excerpt says Windows and AMD were not tested. It does not establish long-context decode speed, broader compatibility, or how other quantizations would behave.
Reader takeaway
Maya’s MiniMax support is a substantial local-serving addition, not a new MiniMax model release. The practical signal is that one inference stack can add a distinct engine and memory-management path to host another model, with reported results on a specific multi-GPU Linux machine. Treat those results as a starting point for your own tests, not a portable performance promise.