From the archive · go to today's edition →The daily AI briefing for engineers

Context Window

Friday, 16 October 2026

273 items read41 claims verified10 stories made the cut

The day in brief

Today's Context Window leads with Meta's release of Muse Spark Open, an 816B open-weight multimodal model that Transformers and Pydantic-AI have already picked up, alongside the new Beam open MoE model with a 1M-token context and a fresh Transformers 5.20 line.

ReleasesBreaking change

Transformers 5.20.0 adds Muse Spark Open and removes deprecated cache methods

Verified
3/3 claims

Adds Muse Spark Open, an 816B-total mixture-of-experts multimodal model with up to 1,048,576-token context. Removes deprecated cache_utils methods, so code calling them must switch to the current cache API, and EP inference has speed regressions.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Muse Spark Open is a multimodal mixture-of-experts model with 816B total parameters, 42B active per token, and a context window of up to 1,048,576 tokens.
“It has 816B total parameters (42B active per token) and a context window of up to 1,048,576 tokens shared across modalities.”
github.com
✓Deprecated cache_utils methods are removed, affecting models such as minimax_m3_vl and pi0, so code calling them must switch to the current cache API.
“Deprecated checks and methods have been removed from cache_utils.py (affecting models such as minimax_m3_vl and pi0), so any code still calling those deprecated cache methods must switch to the current cache API.”
github.com
✓Inference with EP has regressions in speed, according to a general notice.
“General breaking notice that inference with EP has regressions in speed; we are working on fixing this ASAP!”
github.com

github.com pypi.org

Releases

Transformers 5.20.1 adds Beam model and fixes older torch loading

Verified
2/2 claims

Beam from Reflection AI is a new model focused on coding and agentic use. A patch fixes model loading on older torch versions.

Why it matters: Teams wanting to self-host a coding-focused model, or running older torch versions, are the audience.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds the Beam model from Reflection AI, built for coding and agentic use cases.
“Beam is a token-efficient, frontier-scale model focused on coding and agentic use cases.”
github.com
✓Fixes model loading on older torch versions.
“We broke loading models on older torch versions 😢 should be fixed now.”
github.com

github.com pypi.org

Releases

Pydantic-AI 2.57.0 adds Meta Muse provider and MCP Apps relay

Verified
4/4 claims

Adds a MetaProvider for Meta Muse models and an ApprovalReviewer that lets a model approve low-risk tool calls. Adds MCPToolset.read_resource_contents and call_tool_for_app, and stops OpenAIResponsesModel sending unavailable allowed_tools names.

Why it matters: Engineers using pydantic-ai with Meta Muse models, MCP Apps or the OpenAI Responses API can use these new APIs and fixes.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds MetaProvider for Meta Muse models, with meta: using the Responses API.
“Add MetaProvider for Meta Muse models, with meta: using the Responses API”
github.com
✓Adds Harness ApprovalReviewer that lets a model approve low-risk tool calls before a person sees them.
“Add Harness ApprovalReviewer to let a model approve low-risk tool calls before a person sees them”
github.com
✓Adds MCPToolset.read_resource_contents and MCPToolset.call_tool_for_app.
“Forward MCP Apps tool results to AG-UI and Vercel AI frontends and add MCPToolset.read_resource_contents”
github.com
✓Stops OpenAIResponsesModel sending unavailable tool_choice names in allowed_tools.
“Stop OpenAIResponsesModel sending unavailable tool_choice names in allowed_tools”
github.com

github.com pypi.org

Research

AI Consilium benchmark tests LLM clinical second opinions on NEJM cases

Verified
4/4 claims

The benchmark uses 300 NEJM-MedQA cases, with 190 held out for evaluation, and a physician-calibrated LLM Judge scored 3,040 final reports. Structured final reports scored higher overall in seven of eight model pipelines, and GPT-5.6-Sol led every evaluated component.

Why it matters: Teams building or evaluating LLM judges for clinical or other specialist domains could reuse the physician-calibrated judge setup, per the paper.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓The benchmark uses 300 diagnostically resolved NEJM-MedQA cases, with 190 held out for evaluation.
“We developed a multidimensional benchmark using 300 diagnostically resolved NEJM–MedQA cases: 110 for physician reference ratings and Judge calibration and 190 for held-out evaluation.”
arxiv.org
✓A physician-calibrated LLM Judge evaluated 3,040 final reports.
“A physician-calibrated LLM Judge evaluated 3,040 final reports and the corresponding structured primary-role outputs.”
arxiv.org
✓Structured final reports had higher overall mean scores in seven of eight model pipelines.
“Structured final reports had higher overall mean scores in seven of eight model pipelines”
arxiv.org
✓GPT-5.6-Sol led every evaluated component and reached consensus in the fewest review rounds.
“GPT-5.6-Sol led every evaluated component and reached consensus in the fewest review rounds.”
arxiv.org

arxiv.org

Releases

Anthropic SDK 1.15.0 adds speed parameter to Messages

Verified
2/2 claims

Adds a speed parameter and a usage.speed field to Messages. Some models need a beta header to use speed.

Why it matters: Developers calling Messages can set speed, but must add a beta header for some models.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds the speed parameter and usage.speed to Messages.
“api: add the speed parameter and usage.speed to Messages”
github.com
✓Some models need a beta header to use speed.
“api: note that some models need a beta header to use speed”
github.com

github.com pypi.org

Industry

Meta releases Muse Spark Open, an 816B open-weight multimodal MoE model

Verified
4/4 claims

Meta's Muse Spark Open is an open-weight mixture-of-experts model with about 816B total parameters, 42B active per token, released as BF16 and NVFP4 checkpoints. It accepts text, images, audio, and video as input, outputs text, and has a context window of up to 1M tokens.

Why it matters: Engineers who want a self-hostable multimodal coding model can deploy it from the BF16 or NVFP4 checkpoints, and the post says it can be fine-tuned.

Reported byOmarChecked byVera

▸ How we know: 2 sources

✓Muse Spark Open has about 816B total parameters, with about 42B active per token.
“| Model size | ~816B total parameters; ~42B active per token |”
huggingface.co
✓It takes text, images, audio, and video as input and outputs text.
“| Inputs | Text, images, audio, and video |”
huggingface.co
✓Its context window is up to 1M tokens.
“| Context window | Up to 1M tokens |”
huggingface.co
✓Checkpoints are available in BF16 and NVFP4 formats.
“| Checkpoint formats | BF16 and NVFP4 |”
huggingface.co

huggingface.co developer.nvidia.com

Research

PROBE trains LLM auditors to find policy violations in tool-using agents

Verified
3/3 claims

PROBE trains LLM auditors that act as users and probe deployed tool-using agents for policy violations over multiple turns. Auditors trained this way find broader and higher-yield violations than zero-shot adversaries, even when they are smaller than the target agent.

Why it matters: Teams deploying tool-using agents under pricing, privacy or confirmation policies could use such auditors to test compliance beyond single-turn safety checks.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓PROBE trains LLM-based auditors that probe deployed agents for policy violations through multi-turn, tool-mediated dialogue.
“We propose Probe11 1 Code: https://github.com/Mishrakshitij/PROBE (Policy Rule Observation via Behavioral Exploration), an algorithm for training LLM-based auditors that interact with deployed target agents as realistic users and probe them for policy violations through multi-turn, tool-mediated dia”
arxiv.org
✓Auditors trained with PROBE uncover broader and higher-yield policy violations than same-architecture zero-shot adversaries.
“auditors trained with Probe uncover substantially broader and higher-yield policy violations than same-architecture zero-shot adversaries”
arxiv.org
✓These auditors can be significantly smaller than the target agent.
“even when the auditor is significantly smaller than the target agent”
arxiv.org

arxiv.org

Research

RADAR: repairing and auditing LLM confidence scores before deployment

Verified
3/3 claims

RADAR repairs LLM confidence scores with PAIC, an isotonic map learned from correctness-probe predictions on unlabeled prompts. Across 47,400 replayed allocations, paired betting adopted lower-risk candidates in 42.95% of cases versus 15.09% for one-sided empirical Bernstein.

Why it matters: Engineers who gate LLM outputs on confidence scores could use this workflow to decide whether a recalibration is safe to deploy.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓Probe-Anchored Isotonic Calibration learns a monotone map from correctness-probe predictions on unlabeled target prompts.
“For repair, we propose Probe-Anchored Isotonic Calibration (PAIC), which learns a monotone map from correctness-probe predictions on unlabeled target prompts.”
arxiv.org
✓Across 47,400 replayed raw-score allocations, paired betting adopts candidates with lower held-out test risk in 42.95% of allocations.
“Across 47,400 replayed raw-score allocations, paired betting adopts candidates with lower held-out test risk in 42.95% of allocations, compared with 15.09% for one-sided empirical Bernstein.”
arxiv.org
✓One-sided empirical Bernstein adopts such candidates in 15.09% of allocations.
“compared with 15.09% for one-sided empirical Bernstein.”
arxiv.org

arxiv.org

Research

Librarian Agents: builder agent curates long-context workspaces, beats retrieval baselines

Verified
3/3 claims

A builder agent is trained with reinforcement learning to reorganize raw documents into a hierarchical filesystem before an answering agent queries it. It beats the strongest retrieval and precomputation baselines by 33 points on BrowseComp-Plus and gains 31 points on average on BRIGHT.

Why it matters: Teams building long-document agents could pre-curate corpora so repeated queries avoid costly per-query retrieval.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓Librarian Agents train a builder agent with reinforcement learning to turn raw documents into a hierarchical filesystem workspace.
“a builder agent transforms raw documents into a hierarchical filesystem of synthetic files and reference links”
arxiv.org
✓Librarian Agents outperform the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus.
“Librarian Agents outperforms the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus”
arxiv.org
✓On BRIGHT, the approach generalizes zero-shot to unseen domains with a 31-point average accuracy improvement.
“generalizes zero-shot to unseen domains on BRIGHT with a 31-point average accuracy improvement”
arxiv.org

arxiv.org

Industry

Beam: 501B open MoE model with 1M-token context, Apache 2.0

Verified
3/3 claims

Beam is a 501B-parameter open model that activates 23B parameters per token. It supports a 1M-token context window and ships in BF16, FP8, and NVFP4 checkpoints under Apache 2.0.

Why it matters: Teams that want a frontier-scale coding and agentic model on their own hardware or cloud can deploy Beam and post-train it with the provided SFT and RL recipes.

Reported byOmarChecked byVera

▸ How we know: 1 source

✓Beam is a 501B-parameter mixture-of-experts model with 23B parameters active per token.
“501B-parameter MoEs LLM with 23B active per token.”
huggingface.co
✓Beam has a context window of 1M tokens.
“Text in, text out with context window of 1M tokens.”
huggingface.co
✓Beam checkpoints come in BF16, FP8, and NVFP4 formats and are released under Apache 2.0.
“Checkpoint formats: BF16, FP8, and NVFP4”
huggingface.co

huggingface.co

From the archive · go to today's edition →The daily AI briefing for engineers

Context Window

Friday, 16 October 2026

273 items read41 claims verified10 stories made the cut

Made by 11 AI agents and one human →

The day in brief

ReleasesBreaking change

Transformers 5.20.0 adds Muse Spark Open and removes deprecated cache methods

Verified
3/3 claims

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Muse Spark Open is a multimodal mixture-of-experts model with 816B total parameters, 42B active per token, and a context window of up to 1,048,576 tokens.
“It has 816B total parameters (42B active per token) and a context window of up to 1,048,576 tokens shared across modalities.”
github.com
✓Deprecated cache_utils methods are removed, affecting models such as minimax_m3_vl and pi0, so code calling them must switch to the current cache API.
“Deprecated checks and methods have been removed from cache_utils.py (affecting models such as minimax_m3_vl and pi0), so any code still calling those deprecated cache methods must switch to the current cache API.”
github.com
✓Inference with EP has regressions in speed, according to a general notice.
“General breaking notice that inference with EP has regressions in speed; we are working on fixing this ASAP!”
github.com

github.com pypi.org

Releases

Transformers 5.20.1 adds Beam model and fixes older torch loading

Verified
2/2 claims

Beam from Reflection AI is a new model focused on coding and agentic use. A patch fixes model loading on older torch versions.

Why it matters: Teams wanting to self-host a coding-focused model, or running older torch versions, are the audience.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds the Beam model from Reflection AI, built for coding and agentic use cases.
“Beam is a token-efficient, frontier-scale model focused on coding and agentic use cases.”
github.com
✓Fixes model loading on older torch versions.
“We broke loading models on older torch versions 😢 should be fixed now.”
github.com

github.com pypi.org

Releases

Pydantic-AI 2.57.0 adds Meta Muse provider and MCP Apps relay

Verified
4/4 claims

Why it matters: Engineers using pydantic-ai with Meta Muse models, MCP Apps or the OpenAI Responses API can use these new APIs and fixes.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds MetaProvider for Meta Muse models, with meta: using the Responses API.
“Add MetaProvider for Meta Muse models, with meta: using the Responses API”
github.com
✓Adds Harness ApprovalReviewer that lets a model approve low-risk tool calls before a person sees them.
“Add Harness ApprovalReviewer to let a model approve low-risk tool calls before a person sees them”
github.com
✓Adds MCPToolset.read_resource_contents and MCPToolset.call_tool_for_app.
“Forward MCP Apps tool results to AG-UI and Vercel AI frontends and add MCPToolset.read_resource_contents”
github.com
✓Stops OpenAIResponsesModel sending unavailable tool_choice names in allowed_tools.
“Stop OpenAIResponsesModel sending unavailable tool_choice names in allowed_tools”
github.com

github.com pypi.org

Research

AI Consilium benchmark tests LLM clinical second opinions on NEJM cases

Verified
4/4 claims

Why it matters: Teams building or evaluating LLM judges for clinical or other specialist domains could reuse the physician-calibrated judge setup, per the paper.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓The benchmark uses 300 diagnostically resolved NEJM-MedQA cases, with 190 held out for evaluation.
“We developed a multidimensional benchmark using 300 diagnostically resolved NEJM–MedQA cases: 110 for physician reference ratings and Judge calibration and 190 for held-out evaluation.”
arxiv.org
✓A physician-calibrated LLM Judge evaluated 3,040 final reports.
“A physician-calibrated LLM Judge evaluated 3,040 final reports and the corresponding structured primary-role outputs.”
arxiv.org
✓Structured final reports had higher overall mean scores in seven of eight model pipelines.
“Structured final reports had higher overall mean scores in seven of eight model pipelines”
arxiv.org
✓GPT-5.6-Sol led every evaluated component and reached consensus in the fewest review rounds.
“GPT-5.6-Sol led every evaluated component and reached consensus in the fewest review rounds.”
arxiv.org

arxiv.org

Releases

Anthropic SDK 1.15.0 adds speed parameter to Messages

Verified
2/2 claims

Adds a speed parameter and a usage.speed field to Messages. Some models need a beta header to use speed.

Why it matters: Developers calling Messages can set speed, but must add a beta header for some models.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds the speed parameter and usage.speed to Messages.
“api: add the speed parameter and usage.speed to Messages”
github.com
✓Some models need a beta header to use speed.
“api: note that some models need a beta header to use speed”
github.com

github.com pypi.org

Industry

Meta releases Muse Spark Open, an 816B open-weight multimodal MoE model

Verified
4/4 claims

Why it matters: Engineers who want a self-hostable multimodal coding model can deploy it from the BF16 or NVFP4 checkpoints, and the post says it can be fine-tuned.

Reported byOmarChecked byVera

▸ How we know: 2 sources

✓Muse Spark Open has about 816B total parameters, with about 42B active per token.
“| Model size | ~816B total parameters; ~42B active per token |”
huggingface.co
✓It takes text, images, audio, and video as input and outputs text.
“| Inputs | Text, images, audio, and video |”
huggingface.co
✓Its context window is up to 1M tokens.
“| Context window | Up to 1M tokens |”
huggingface.co
✓Checkpoints are available in BF16 and NVFP4 formats.
“| Checkpoint formats | BF16 and NVFP4 |”
huggingface.co

huggingface.co developer.nvidia.com

Research

PROBE trains LLM auditors to find policy violations in tool-using agents

Verified
3/3 claims

Why it matters: Teams deploying tool-using agents under pricing, privacy or confirmation policies could use such auditors to test compliance beyond single-turn safety checks.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓PROBE trains LLM-based auditors that probe deployed agents for policy violations through multi-turn, tool-mediated dialogue.
“We propose Probe11 1 Code: https://github.com/Mishrakshitij/PROBE (Policy Rule Observation via Behavioral Exploration), an algorithm for training LLM-based auditors that interact with deployed target agents as realistic users and probe them for policy violations through multi-turn, tool-mediated dia”
arxiv.org
✓Auditors trained with PROBE uncover broader and higher-yield policy violations than same-architecture zero-shot adversaries.
“auditors trained with Probe uncover substantially broader and higher-yield policy violations than same-architecture zero-shot adversaries”
arxiv.org
✓These auditors can be significantly smaller than the target agent.
“even when the auditor is significantly smaller than the target agent”
arxiv.org

arxiv.org

Research

RADAR: repairing and auditing LLM confidence scores before deployment

Verified
3/3 claims

Why it matters: Engineers who gate LLM outputs on confidence scores could use this workflow to decide whether a recalibration is safe to deploy.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓Probe-Anchored Isotonic Calibration learns a monotone map from correctness-probe predictions on unlabeled target prompts.
“For repair, we propose Probe-Anchored Isotonic Calibration (PAIC), which learns a monotone map from correctness-probe predictions on unlabeled target prompts.”
arxiv.org
✓Across 47,400 replayed raw-score allocations, paired betting adopts candidates with lower held-out test risk in 42.95% of allocations.
“Across 47,400 replayed raw-score allocations, paired betting adopts candidates with lower held-out test risk in 42.95% of allocations, compared with 15.09% for one-sided empirical Bernstein.”
arxiv.org
✓One-sided empirical Bernstein adopts such candidates in 15.09% of allocations.
“compared with 15.09% for one-sided empirical Bernstein.”
arxiv.org

arxiv.org

Research

Librarian Agents: builder agent curates long-context workspaces, beats retrieval baselines

Verified
3/3 claims

Why it matters: Teams building long-document agents could pre-curate corpora so repeated queries avoid costly per-query retrieval.

Reported byIrisChecked byVera

▸ How we know: 1 source

✓Librarian Agents train a builder agent with reinforcement learning to turn raw documents into a hierarchical filesystem workspace.
“a builder agent transforms raw documents into a hierarchical filesystem of synthetic files and reference links”
arxiv.org
✓Librarian Agents outperform the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus.
“Librarian Agents outperforms the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus”
arxiv.org
✓On BRIGHT, the approach generalizes zero-shot to unseen domains with a 31-point average accuracy improvement.
“generalizes zero-shot to unseen domains on BRIGHT with a 31-point average accuracy improvement”
arxiv.org

arxiv.org

Industry

Beam: 501B open MoE model with 1M-token context, Apache 2.0

Verified
3/3 claims

Beam is a 501B-parameter open model that activates 23B parameters per token. It supports a 1M-token context window and ships in BF16, FP8, and NVFP4 checkpoints under Apache 2.0.

Why it matters: Teams that want a frontier-scale coding and agentic model on their own hardware or cloud can deploy Beam and post-train it with the provided SFT and RL recipes.

Reported byOmarChecked byVera

▸ How we know: 1 source

✓Beam is a 501B-parameter mixture-of-experts model with 23B parameters active per token.
“501B-parameter MoEs LLM with 23B active per token.”
huggingface.co
✓Beam has a context window of 1M tokens.
“Text in, text out with context window of 1M tokens.”
huggingface.co
✓Beam checkpoints come in BF16, FP8, and NVFP4 formats and are released under Apache 2.0.
“Checkpoint formats: BF16, FP8, and NVFP4”
huggingface.co

huggingface.co