Large Language Models

Today's Stories

1. 1

   Nvidia RTX Spark: 1 PFLOP, 120B LLM, From $1,799 [2026]

   Nvidia RTX Spark: 1 PFLOP, 120B LLM, From $1,799 [2026]

2. 2

   DeepSeek R1 becomes #2 open-weights reasoning model

   DeepSeek R1, an open-weights reasoning model, has reached the #2 position in its category. The model processes 1M tokens (an increase from 128K in V3.2) and operates on 27% of FLOPs compared with DeepSeek-V3.2. Open-weights models like DeepSeek R1 let any organization run powerful AI reasoning without relying on proprietary platforms. This shift may reshape how businesses approach complex problem-solving tasks, since they can now deploy advanced reasoning capabilities on their own infrastructure.

   The jump in context window from 128K to 1M tokens represents a significant expansion in how much information the model can process at once, potentially enabling new use cases in document analysis and longer-form reasoning tasks.

3. 3

   NEC's cotomi LLM runs inference in confidential computing, Acompany partnership

   NEC's domestically developed language model cotomi successfully executed inference in a confidential computing environment through a proof-of-concept with Acompany. Confidential computing allows data to be processed while remaining encrypted and hidden from the underlying system. The ability to run AI inference on sensitive data without exposing it—even to infrastructure operators—addresses a major concern for businesses handling confidential information. This approach may allow enterprises to adopt AI capabilities while protecting proprietary or regulated data.

   The companies demonstrated the capability through their joint project; further details on timeline, deployment scope, or commercial availability are not yet disclosed in the announcement.

4. 4

   Claude Opus 4.7 tops new code-recreation benchmark at 56% solve rate

   Epoch AI released MirrorCode, a benchmark that tests whether AI models can rebuild complete programs from scratch without seeing the original code. Claude Opus 4.7 leads with a 56 percent solve rate and rebuilt a 16,000-line toolkit in just 14 hours, but all models tested still fail on the most complex tasks. The benchmark reveals both progress and limits in AI coding ability. Strong performance on medium-difficulty reconstruction suggests these models are becoming more capable at reverse-engineering software logic, but consistent failure on harder problems shows they cannot yet handle the most intricate programming challenges.

   The benchmark includes tasks expensive enough that one required $2,600 to run and 19 days of continuous compute time, suggesting some code-recreation problems push AI systems to their practical limits.

5. 5

   OpenAI launches GPT-5.6 Sol, rivals Claude in coding under US government access limits

   OpenAI unveiled GPT-5.6 Sol, a new flagship model competing with Anthropic's Claude Mythos. Access is limited to select partners through the API and Codex at the US government's direction. The model comes in three tiers—Sol (flagship), Terra (matching GPT-5.5 performance at half the cost), and Luna (budget option)—plus "max" and "ultra" reasoning modes. On Terminal-Bench 2.1, Sol scores 88.8 percent in agentic coding (AI agents that independently handle complex tasks), edging Claude Mythos 5's 88 percent. Sol Ultra reaches 91.9 percent. On cybersecurity benchmarks, Sol matches Claude's performance while using roughly a third of the output tokens, suggesting lower effective costs per task despite broader inflation in AI pricing. OpenAI has publicly objected to government-mandated access restrictions, calling them unsustainable.

   Sol pricing is $5 per million input tokens and $30 per million output tokens; Terra costs $2.50 and $15; Luna costs $1 and $6. In July, Sol launches on Cerebras at up to 750 tokens per second. OpenAI has also revamped prompt caching with a guaranteed 30-minute minimum lifetime and 90 percent discounts on cache reads.

6. 6

   White House to decide frontier AI model access on ad hoc basis

   The White House is establishing a new policy under which it will individually and opaquely decide who can access frontier AI models on a case-by-case basis, rather than using a formal, predictable set of procedures. This approach leaves the criteria for access unclear and subject to discretionary judgment rather than transparent rules. The author notes that groundwork for a formal process was not laid over preceding years, leaving the country without a predictable framework for managing frontier AI deployment.

   The author expresses hope that this ad hoc system will eventually transition into a formal and predictable set of procedures, though the current policy as described does not yet provide that clarity.