This frozen page shows Augur's claims and source links for one sent dispatch. Stored spot-checks appear only where the frozen edition supports them; absence is not presented as verification.
Apple's research group has now moved that ceiling a long way: it ran Meta's Muse Spark Open, an 816-billion-parameter model (the size of a lab's top tier), across four Mac Studios with M5 Ultra chips and got more than 36 tokens per second, the word fragments models emit Apple Machine Learning Research.
Assertion status: No spot-check verdict is published for this assertion.
Apple Research reports that running the Muse Spark Open model (816B parameters) on a cluster of four Mac Studios with M5 Ultra chips generates over 36 tokens per second.
Claim 80097Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio achieves up to 4.6 times faster time-to-first-token compared to the M3 Ultra Mac Studio due to GPU Neural Accelerators.
Claim 80098Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio provides 1.6 times faster generation speed (tokens per second) than the M3 Ultra Mac Studio due to increased unified memory bandwidth.
Claim 80099Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 2
Muse Spark Open is a mixture-of-experts model, meaning only about 42 billion of its 816 billion parameters do work on any given token, so the machines must hold everything in memory but compute only a slice per word, and the M5 Ultra Mac Studio can be configured with up to 512 GB of memory, the kind of room a four-way split of a model this size needs Hugging Face Blog Nate Jones.
Assertion status: No spot-check verdict is published for this assertion.
The base Mac mini is equipped with the new M6 chip, while more powerful models like the Mac mini, Mac Studio, and Mac Pro variants use the M5, M5 Pro, M5 Max, and M5 Ultra chips.
Claim 57658Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 3
Apple's researchers tied the four boxes together with MLX (Apple's software for running AI models on its own chips) and RDMA (a way for one machine to read another's memory directly), a capability the exo project says Apple promotes on its product pages as low-latency RDMA over Thunderbolt 5 Apple Machine Learning Research Latent Space.
Assertion status: No spot-check verdict is published for this assertion.
The exo project claims that Apple featured its technology on new M5 Ultra Mac Studio and M6/M5 Pro Mac Mini pages, emphasizing low-latency RDMA over Thunderbolt 5.
Claim 52403Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research reports that running the Muse Spark Open model (816B parameters) on a cluster of four Mac Studios with M5 Ultra chips generates over 36 tokens per second.
Claim 80097Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio achieves up to 4.6 times faster time-to-first-token compared to the M3 Ultra Mac Studio due to GPU Neural Accelerators.
Claim 80098Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio provides 1.6 times faster generation speed (tokens per second) than the M3 Ultra Mac Studio due to increased unified memory bandwidth.
Claim 80099Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 4
The hardware then turns that fit into usable speed: Apple attributes an up to 4.6 times shorter wait before the first word to new accelerator units on the graphics chip, and 1.6 times faster generation to memory bandwidth, both measured against the older M3 Ultra Apple Machine Learning Research.
Assertion status: No spot-check verdict is published for this assertion.
Apple Research reports that running the Muse Spark Open model (816B parameters) on a cluster of four Mac Studios with M5 Ultra chips generates over 36 tokens per second.
Claim 80097Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio achieves up to 4.6 times faster time-to-first-token compared to the M3 Ultra Mac Studio due to GPU Neural Accelerators.
Claim 80098Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio provides 1.6 times faster generation speed (tokens per second) than the M3 Ultra Mac Studio due to increased unified memory bandwidth.
Claim 80099Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 5
Priced against hosted Muse Spark, the closed sibling, at $4.25 per million output tokens, that would displace roughly $13 a day in output fees, and at Meta's $0.20 contributor tier the figure drops well under a dollar Big Technology Simon Willison's Weblog.
Assertion status: No spot-check verdict is published for this assertion.
Muse Spark 1.2 is available under two model IDs: muse-spark-1.2 priced at $1.25/million input tokens and $4.25/million output tokens, and muse-spark-1.2-contributor priced at $0.10/million input tokens and $0.20/million output tokens.
Claim 42896Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 6
The supplied pricing evidence gives no hosted price for Muse Spark Open itself, even though the same open weights can now be rented through Amazon SageMaker JumpStart, and nothing here prices the four-Mac cluster, so any saving has to be set against your own hardware quote AWS Machine Learning Blog.
Assertion status: No spot-check verdict is published for this assertion.
Muse Spark Open uses a mixture-of-experts architecture with 816 billion total parameters and 42 billion active parameters per token.
Claim 79703Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 7
That rental option actually sharpens the real case, which is control: a hosted deployment, SageMaker included, still sends your documents off site, while a cluster you own means no rate limits, no surprise repricing by a vendor, and data that never leaves the building, which is why confidentiality-bound professional services firms are already buying Mac Minis to run local models Nate Jones.
Assertion status: No spot-check verdict is published for this assertion.
Professional services firms with high confidentiality requirements are increasingly purchasing Mac Minis to run local AI models.
Claim 15124Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 8
Reflection AI launched Beam on October 5 as a US-trained open-weight (downloadable model weights) answer to Chinese models, and this week the plumbing caught up: Hugging Face published its welcome post and shipped native Beam support in Transformers 5.20.1, a patch that also repairs a 5.20.0 break in loading models on older versions of PyTorch, the underlying math library TechCrunch AI Hugging Face Blog Transformers Releases.
Assertion status: No spot-check verdict is published for this assertion.
Reflection AI unveiled its first frontier open-weight AI model named Beam on October 5, 2026.
Claim 73152Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Reflection AI claims its Beam model matches the performance of leading Chinese open models on advanced reasoning benchmarks at significantly lower costs.
Claim 73153Label: opinionProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Reflection AI positions Beam as a direct rival to closed labs like Anthropic and OpenAI, Chinese open models, and Western players like Mistral, Meta, and Cohere.
Claim 73158Label: opinionProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 9
Beam fires 23 billion of its 501 billion parameters per token, and its 1 million token context (how much text the model can read at once) comes from a mix of layers, 43 that look only at nearby text and nine that scan the whole document Hugging Face Blog.
Assertion status: No spot-check verdict is published for this assertion.
Reflection released an open-source 501-billion-parameter Mixture-of-Experts language model named Beam.
Claim 80333Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 10
The score picture is disputed: Reflection says Beam scores as well as the best Chinese open models on hard reasoning tests while costing less to run, while Import AI's read is that it trails the Chinese model GLM-5.3 on benchmarks and leans on a token-efficiency claim, meaning it reaches the same answer while generating fewer word fragments TechCrunch AI Import AI.
Assertion status: No spot-check verdict is published for this assertion.
Reflection AI unveiled its first frontier open-weight AI model named Beam on October 5, 2026.
Claim 73152Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Reflection AI claims its Beam model matches the performance of leading Chinese open models on advanced reasoning benchmarks at significantly lower costs.
Claim 73153Label: opinionProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Reflection AI positions Beam as a direct rival to closed labs like Anthropic and OpenAI, Chinese open models, and Western players like Mistral, Meta, and Cohere.
Claim 73158Label: opinionProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Reflection announced Beam, a 501 billion parameter MoE LLM, which is behind Chinese models like GLM-5.3 in benchmark scores but claims higher token efficiency.
Claim 76772Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 11
Google Research has built VERID, a digital twin (a live software copy of the physical lab) on the open-source PyLabRobot, so an AI agent can dry-run and safety-check a procedure before a robot touches a pipette, and in aptamer experiments (short DNA strands selected to bind a target) confirmed hit rates, the share of designed candidates that actually bound the target, reached 33.5% and 42.1% after the agents spotted and corrected their own wet-lab failures Google Research Blog.
Assertion status: No spot-check verdict is published for this assertion.
Google Research is integrating AI agents with physical wet-lab robotics to enable closed-loop molecular design cycles for discovering protein binders with single-digit nanomolar affinity for cancer targets.
Claim 80512Label: forecastProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
The VERID (VERified Infrastructure for Discovery Laboratories) framework is a semantic digital twin built on the open-source PyLabRobot that provides live 3D state tracking, virtual dry runs, hardware safety checks, and synchronized video logs to bridge the gap between AI reasoning and physical lab operations.
Claim 80513Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
In aptamer discovery experiments targeting myoglobin and α-thrombin, AI-driven design cycles raised confirmed on-target hit rates to 33.5% and 42.1% respectively, after agents autonomously identified and corrected wet-lab failure modes such as primer overhang disruption and non-specific binding.
Claim 80514Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 13
- Independent long-context results for Reflection's Beam would test the token-efficiency claim against the disputed benchmark picture. Import AI
Assertion status: No spot-check verdict is published for this assertion.
Reflection announced Beam, a 501 billion parameter MoE LLM, which is behind Chinese models like GLM-5.3 in benchmark scores but claims higher token efficiency.
Claim 76772Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 14
- Google Research has promised closed-loop discovery of tight-binding proteins; a published binder with a named target is the proof point. Google Research Blog
Assertion status: No spot-check verdict is published for this assertion.
Google Research is integrating AI agents with physical wet-lab robotics to enable closed-loop molecular design cycles for discovering protein binders with single-digit nanomolar affinity for cancer targets.
Claim 80512Label: forecastProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
The VERID (VERified Infrastructure for Discovery Laboratories) framework is a semantic digital twin built on the open-source PyLabRobot that provides live 3D state tracking, virtual dry runs, hardware safety checks, and synchronized video logs to bridge the gap between AI reasoning and physical lab operations.
Claim 80513Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
In aptamer discovery experiments targeting myoglobin and α-thrombin, AI-driven design cycles raised confirmed on-target hit rates to 33.5% and 42.1% respectively, after agents autonomously identified and corrected wet-lab failure modes such as primer overhang disruption and non-specific binding.
Claim 80514Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 15
- The supplied pricing evidence gives no hosted price for Muse Spark Open; a listing at or below Muse Spark's $4.25 output rate, on SageMaker or elsewhere, would shrink the local cost case further. Big Technology AWS Machine Learning Blog
Assertion status: No spot-check verdict is published for this assertion.
Meta's Muse Spark costs $1.25 per million input tokens and $4.25 per million output tokens.
Claim 67854Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 16
- Black Forest Labs has now shipped FLUX 3 Image Klein, the 5.6B open-weight model, after promising an open-weight variant at the October 1 launch; independent tests of its 4K generation and editing claims are the next check. Latent Space Black Forest Labs
Assertion status: No spot-check verdict is published for this assertion.
Black Forest Labs launched FLUX 3 Image, a model capable of native generation up to 4K, up to ten reference images, bounding-box layout control, and targeted multi-turn editing.
Claim 73418Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Black Forest Labs announced commercial weights for FLUX 3 Image are available, with an open-weight variant promised in coming weeks, and hosted access through fal and Krea.
Claim 73420Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Black Forest Labs released FLUX 3 Image Klein, a 5.6B parameter open-weight model for text-to-image generation, image editing, and multi-reference generation.
Claim 80117Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
FLUX 3 Image Klein shares core capabilities with the full FLUX 3 Image model, including native 4K generation, pixel-perfect editing, and bounding box controls.
Claim 80118Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
FLUX 3 Image Klein builds on FLUX 3's shared learnings across images, video, and audio, using spatial detail from images, motion from video, and event information from audio.
Claim 80119Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 17
- Hugging Face's follow-on Transformers patches should show whether the older-PyTorch loading break in 5.20.0 was fully closed by 5.20.1. Transformers Releases
Assertion status: No spot-check verdict is published for this assertion.
The Hugging Face Transformers library version 5.20.1 includes a new model named Beam from Reflection AI.
Claim 80388Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 18
Apple's 36 tokens per second is a vendor-reported figure, and the figures available here do not say whether it covers one conversation or many served at once, which changes the cost arithmetic above Apple Machine Learning Research.
Assertion status: No spot-check verdict is published for this assertion.
Apple Research reports that running the Muse Spark Open model (816B parameters) on a cluster of four Mac Studios with M5 Ultra chips generates over 36 tokens per second.
Claim 80097Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio achieves up to 4.6 times faster time-to-first-token compared to the M3 Ultra Mac Studio due to GPU Neural Accelerators.
Claim 80098Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Apple Research states that the M5 Ultra Mac Studio provides 1.6 times faster generation speed (tokens per second) than the M3 Ultra Mac Studio due to increased unified memory bandwidth.
Claim 80099Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 19
The 512 GB memory figure is the top M5 Ultra configuration, not a stated spec of Apple's test machines Nate Jones.
Assertion status: No spot-check verdict is published for this assertion.
The base Mac mini is equipped with the new M6 chip, while more powerful models like the Mac mini, Mac Studio, and Mac Pro variants use the M5, M5 Pro, M5 Max, and M5 Ultra chips.
Claim 57658Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 20
The $4.25 and $0.20 output prices are for hosted Muse Spark tiers, not the Open variant, and the daily-saving figures are hypotheticals built on them with no cluster price to set against; input fees ($1.25 per million on hosted Muse Spark) are left out because the evidence gives no figure for how fast the cluster processes input Big Technology Simon Willison's Weblog.
Assertion status: No spot-check verdict is published for this assertion.
Muse Spark 1.2 is available under two model IDs: muse-spark-1.2 priced at $1.25/million input tokens and $4.25/million output tokens, and muse-spark-1.2-contributor priced at $0.10/million input tokens and $0.20/million output tokens.
Claim 42896Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
Assertion 21
Google's hit rates come from its own writeup, the hit rate before the agents' corrections is not in the evidence here, and the binder goal is a forecast Google Research Blog.
Assertion status: No spot-check verdict is published for this assertion.
Google Research is integrating AI agents with physical wet-lab robotics to enable closed-loop molecular design cycles for discovering protein binders with single-digit nanomolar affinity for cancer targets.
Claim 80512Label: forecastProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
The VERID (VERified Infrastructure for Discovery Laboratories) framework is a semantic digital twin built on the open-source PyLabRobot that provides live 3D state tracking, virtual dry runs, hardware safety checks, and synchronized video logs to bridge the gap between AI reasoning and physical lab operations.
Claim 80513Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded
No stored spot-check names this claim in this edition.
In aptamer discovery experiments targeting myoglobin and α-thrombin, AI-driven design cycles raised confirmed on-target hit rates to 33.5% and 42.1% respectively, after agents autonomously identified and corrected wet-lab failure modes such as primer overhang disruption and non-specific binding.
Claim 80514Label: factProvenance: primaryEligible for corroboration · eligibility does not establish truthRecorded