MiniMax rolls out M3.1 Flash for coding and agents with 1M context
Released
MiniMax formally rolled out M3.1 Flash through its API, M Plan and MiniMax Code, following the earlier Space Bunny test. It supports text, image and video inputs, a 1M-token context and five thinking-depth settings. Posted API rates are $0.15 per million input tokens and $0.60 per million output tokens.
Why it matters
Developers gain another low-cost option for long-context coding, computer use and multi-step agent workflows, with OpenAI- and Anthropic-compatible APIs.
Caveat:Status: confirmed_release. Verification: primary. Speed, task-cost reductions and benchmark comparisons are MiniMax claims, not independently reproduced results. The launch post says weights will be released later, despite open-weight wording elsewhere in its docs. Thinking cannot be disabled and defaults to max effort. Timing: date-only; cutoff overlap possible.
02Model releases
ReleasedPrimary source
Inception releases Mercury 3, extending diffusion language models to 1M context
Released
Inception launched Mercury 3 through its API and named inference partners, with a 1M-token context versus 260K for Mercury 2.5. The company reports 1,068 tokens per second on NVIDIA GPUs and comparisons with speed-focused frontier models. Launch pricing is $0.10 per million input tokens and $0.35 per million output tokens.
Why it matters
A larger diffusion language model offers another approach to reducing waiting time in interactive applications and agents that make many sequential model calls.
Caveat:Status: confirmed_release. Verification: primary. Throughput, scaling and quality comparisons are vendor-reported and depend on the workload and serving setup. Hosted API availability is verified from the launch and quick-start documentation; no downloadable Mercury 3 weights were established. Launch pricing is discounted. Timing: date-only; cutoff overlap possible.
03Local and on-device AI
ReleasedPrimary source
SGLang 0.5.22 adds faster loading, Apple Silicon support and a decoding-state fix
Released
SGLang released 0.5.22 with an InstantTensor loader, standard PyTorch execution on Apple Silicon alongside MLX, expanded model support and a fix for incorrect recurrent states when speculative decoding and structured output are combined. Its release notes report faster DeepSeek-V4.1 serving on H200 and support for AMD MI350X and MI355X.
Why it matters
Self-hosters get broader hardware choices and a correctness fix, alongside serving improvements that can shorten model startup and reduce infrastructure costs.
Caveat:Status: confirmed_release. Verification: primary. The advertised 10x loading and up-to-2.3x H200 throughput improvements are project measurements, not universal multipliers. Hardware, model and configuration matter. The release requires Python 3.11 or newer. Timing: date-only; cutoff overlap possible.
04Open-weight projects
ReleasedPrimary source
Inco releases a DFlash 2 drafter for Qwen3.8 Flash Next
Released
Inco AI published a roughly 0.6B-parameter auxiliary checkpoint that drafts blocks of tokens for Qwen3.8 Flash Next to verify. Weights, configuration and an SGLang integration patch are available. Inco reports higher serving throughput than the target's built-in multi-token prediction in its tested configurations.
Why it matters
Serving teams can test speculative decoding on another Qwen target without retraining the main model, potentially improving latency or aggregate throughput.
Caveat:Status: confirmed_release. Verification: primary. This is a new target-specific drafter, not a new standalone chat model or the original DFlash 2 launch. Measurements use one GB300 and a specific FP8 target; gains need independent workload testing. The drafter lists Apache 2.0, while the target has separate terms. The documented path needs a pinned SGLang revision and patch. Timing: date-only; cutoff overlap possible.
05Developer tools
AnnouncedPrimary source
AWS and PyTorch announce TorchNeuron, with public beta planned for October 20
Announced
AWS and PyTorch introduced TorchNeuron, a native backend for Trainium2 and Trainium3 supporting eager execution, torch.compile, distributed APIs and standard profiling tools. The announcement describes moving compatible code to the neuron device and plans an Apache 2.0 open-source public beta on October 20.
Why it matters
A native backend could reduce the model and library changes needed to evaluate Trainium as an alternative accelerator for training and inference.
Caveat:Status: announcement. Verification: primary. The new public beta and source release are scheduled for October 20, so this briefing does not describe them as already shipped. Partner performance examples are workload-specific claims. Experimental adaptive eager mode and privately tested features have separate availability limits. Timing: date-only; cutoff overlap possible.
