Inception says Mercury 3 reaches 1,068 tokens/s and supports a 1-million-token context window, with launch pricing of $0.10/$0.35 per million input/output tokens. MiniMax M3.1 Flash also emphasizes low pricing, but AA measured a 93% hallucination rate, linked to the model attempting to answer almost every question. DeepSeek open-sourced its offline inference framework DeepInfer, while vLLM v0.32.0 increases requests per second with multiple API workers by 43% to 85%. Google open-sourced Harness², which improves the agent harness while keeping model weights frozen, raising Gemini 3.5 Flash's scores on two task benchmarks by 8.4 and 7.5 points, respectively.
Mercury 3, M3.1 Flash, and the latest DeepSeek update put price, speed, and answer quality on the same model-selection checklist.
Inception launched the diffusion language model Mercury 3, claiming a speed of 1,068 tokens/s on NVIDIA GPUs. It has more than 40 times as many parameters as Mercury 1 while maintaining roughly the same speed, and its context window has grown from 260,000 to 1 million tokens. The model is available through Inception's API and platforms including Baseten, with launch prices of $0.10/$0.35 per million input/output tokens—half the regular rates.
Read original →MiniMax launched M3.1 Flash, priced at $0.15 per million input tokens and $0.60 per million output tokens. The company says it is three times as fast as M3 at one-third the cost, with weights scheduled for release on October 23. However, AA's independent evaluation measured a 93% hallucination rate, linked to the model attempting to answer almost every question.
Read original →DeepSeek V4.1 Flash 1017 gained 6 points on the AA Intelligence Index to reach 45, tying GLM-5.3 for second among open-weight models, at a cost of $0.26 per task. Its GDPval-AA Elo rose by 146 to 1746, and its CritPt score increased from 14% to 24%. On AA-Omniscience, its non-hallucination rate rose from 4% to 33%.
Read original →🔍 Analysis: Mercury 3's launch prices for input and output tokens are lower than M3.1 Flash's, but this is a half-price promotion, not necessarily a lasting cost advantage. M3.1 Flash's 93% hallucination rate and the latest DeepSeek version's 33% non-hallucination rate come from different tests and cannot be used to rank them directly. Together, however, they highlight the need for developers choosing low-cost models to examine both refusal behavior and factual reliability, rather than deciding solely on token prices. DeepSeek's cost per task is another way of measuring costs and should not be compared directly with prices per million tokens.
DeepInfer opens up some components used in production serving, while vLLM continues to optimize request handling across multiple API workers.
DeepSeek open-sourced DeepInfer, the inference framework behind its large-scale production serving, under the Apache 2.0 license. This release provides offline inference capabilities, including optimized operators, asynchronous scheduling, and a decoding acceleration mechanism that uses draft tokens and confidence scores to schedule verification. Prefill–decode disaggregation and load balancing across data-parallel replicas, both needed for a complete inference serving system, are not included in this release.
Read original →vLLM released v0.32.0, with 814 commits and a 43% to 85% increase in requests per second with multiple API workers. The release adds the /v1/systemone structured decision endpoint and expands low-precision KV support across hardware, including NVFP4 and 4-bit KV on ROCm. Model Runner V2 also adds a GPU n-gram speculator.
Read original →🔍 Analysis: Both DeepInfer and vLLM introduce speculative acceleration during decoding, but the releases differ in scope: DeepInfer's release covers offline inference, while vLLM's also addresses request throughput with multiple API workers. For deployment teams, DeepInfer's Apache 2.0 license makes component reuse more straightforward, but it does not provide DeepSeek's complete serving stack. Online deployment still requires prefill–decode disaggregation and load balancing across replicas; faster offline inference should not be equated with greater serving capacity.
Muse Glimmer packs capabilities into a model that can run locally, while Harness² seeks gains outside the weights.
Meta released a technical report on Muse Glimmer, a 30-billion-parameter open-weight model designed for agent tasks. It can run at practical speeds with 4-bit precision on a Mac or PC equipped with a single consumer GPU. Training uses full-KL co-distillation with logits from Muse Spark's training process, switching to cross-entropy for the final 5%, and includes weeks of simulated “digital life.”
Read original →Google Research open-sourced Harness², which lets agents modify their own prompts, skills, hooks, and memory while model weights remain frozen. It improves the harness by proposing edits, comparing against baseline trajectories, and retaining lessons learned, without a scorer. Gemini 3.5 Flash gained 8.4 and 7.5 points on WorkBuddy-Bench and JobBench, respectively.
Read original →🔍 Analysis: Muse Glimmer develops agent capabilities through distillation and weeks of simulated “digital life,” while Harness² freezes model weights and instead modifies prompts, skills, hooks, and memory. These are improvements at different layers: the former affects base-model selection and local hardware requirements, while the latter shows that task performance is not determined by model weights alone. For developers, Harness²'s gains on two benchmarks highlight the need to evaluate models and harnesses separately when comparing agents.
Voxtral 3 connects speech processing with tool calls, offering both EU-hosted and self-hosted options.
Mistral launched the Voxtral 3 voice models and agent SDK, with transcription support for more than 50 languages. The SDK combines transcription, reasoning, tool calls, and speech generation in a single real-time pipeline, enabling multi-step tasks during calls. The Mini transcription model has 4 billion parameters and costs $0.0035 per minute. The suite is available through an EU-hosted API or self-hosting, giving enterprises control over where audio is processed and whether conversations are retained.
Read original →Anthropic's OSS Scanner reduces the wait for reports, but scan results receive no human review.
Anthropic launched OSS Scanner, offering free, recurring vulnerability scans to approved critical open-source projects. Unlike the existing process, in which reports are sent only after human review, the service sends model-generated scan results directly to maintainers to reduce waiting times. Reports are not reviewed by humans. Projects must apply, and new projects or forks without distinct features are generally not accepted.
Read original →