Entries in Category “Uncategorized”

All 13 open-weight models with 1 trillion or more parameters since 2024 come from Chinese labs

by

Log-scale scatter of total parameters (billions) against release date from 2024 to October 2026 for 116 open-weight language models of at least 100 billion parameters, split into China and Rest of world; only Chinese models cross the dashed 1-trillion line, starting in July 2025 and reaching 2.8 trillion in July 2026, while the largest non-Chinese model sits just under 1 trillion.

Of the 116 open-weight language models with at least 100 billion parameters that Epoch AI records since 2024-01-01, 70 come from Chinese developers, and all 13 that reach 1 trillion parameters or more do, from Moonshot’s Kimi K2 in July 2025 to Kimi K3 (2.8 trillion) and Alibaba’s Qwen3.8 (2.4 trillion) in mid-2026. The largest open-weight model has grown about fourfold since DeepSeek-V3’s 671 billion parameters in December 2024, while the biggest from a developer outside China is Thinking Machines’ Inkling at 975 billion. Moonshot’s 3-trillion-class Kimi K3.1 Pro, covered by SI News, would extend that lead when its weights arrive on November 4, though size alone has not closed the capability gap to closed models.

Sources & Notes

  • Original story: Moonshot Launches Kimi K3.1 Pro Open Model | UK Funds £100M AI Biosecurity Lab | Anthropic Details Mythos 5 GitHub Deception
  • Data sources: Epoch AI (2026). Data on AI Models (all models CSV)
  • Other references: Epoch AI (2026). Data on AI Models; Moonshot AI (2026). Kimi K3 model card; Qwen Team, Alibaba Group (2026). Qwen3.8-2.4T-A95B model card
  • Notes: Each marker is one model in Epoch AI’s database whose domain includes language, whose accessibility is any open-weights category (unrestricted, restricted use or non-commercial), whose recorded total parameter count is at least 100 billion, and whose Epoch publication date falls between 2024-01-01 and 2026-10-25; duplicate names were dropped. Mixture-of-experts models are plotted at total, not active, parameters (Kimi K3 activates 104 billion of 2.8 trillion), and the dashed line marks 1 trillion. Country is Epoch’s developer-organization country, with one model lacking a country (Cogito v2.1) counted in Rest of world; Epoch publication dates can precede weight release (Kimi K3 was announced 2026-07-16 and its weights published 2026-07-27). Kimi K3.1 Pro is not plotted because Epoch lists it as API access with no parameter count. Data downloaded 2026-10-25; Kimi K3 and Qwen3.8 parameter counts were checked against their model cards.

• 0 Comments

Best open-weight models trail the closed frontier by 11 ECI points, nearly double the 2024–2025 average

by

Line scatter of the monthly gap in Epoch Capabilities Index points between the best closed-weight and best open-weight AI models from 2024-01-01 to 2026-10-01; the gap mostly ranges from 1 to 9 points in 2024 and 2025 around a 5.9-point average, then rises to 10 to 12 points in 2026, dipping to 6.1 in August before reaching 11.0 in October.

On the Epoch Capabilities Index (ECI), the best closed-weight model (Claude Opus 5.5, 168.8) led the best open-weight model (Moonshot’s Kimi K3, 157.9) by 11.0 points on 2026-10-01, compared with an average gap of 5.9 points across 2024 and 2025 and 10.1 points so far in 2026. Kimi K3’s July release briefly cut the gap to 6.1 points before Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 reopened it in September, so open-weight models are improving quickly but the closed frontier has moved faster this year. That is the bar facing Kimi K3.1 Pro, a post-trained successor to K3 whose strong self-reported coding and agent benchmarks SI News calls a notable new challenge to proprietary frontier models, with weights promised by November 4 and no ECI score yet.

Sources & Notes

  • Original story: Kimi K3.1 Pro launches as a 3T-class open-weight model | OpenAI moves frontier training to offline environments
  • Data sources: Epoch AI (2026). Epoch Capabilities Index (ECI) model scores
  • Other references: Epoch AI (2026). Epoch Capabilities Index; Edwards, J., & Emberson, L (2026). Open models lag state-of-the-art closed models by 4 months; Moonshot AI (2026). Introducing Kimi K3.1 Pro
  • Notes: Each point is the gap, on the first day of each month from 2024-01-01 to 2026-10-01, between the highest central ECI estimate among models in Epoch’s Closed weights group and the highest among models in its Open weights group (which includes restricted-use and non-commercial licenses; Kimi K3 is non-commercial) released on or before that date. Models Epoch classifies as Other (unknown or limited access, such as Gemini 4 Argon at 166.5) are excluded, and the gap was unchanged at 11.0 points on 2026-10-24. Data were downloaded on 2026-10-24 from the CSV Epoch updated that day; the confidence intervals for the current leaders span roughly 5 to 9 points and shift when Epoch refits the index, which is why Epoch’s May 2026 analysis reported an average gap of about 8 points since January 2026 while the current fit gives about 10.5 points for 2026-01-01 to 2026-05-29. Kimi K3.1 Pro’s benchmark results are self-reported by Moonshot AI and not yet in the ECI.

• 1 Comment

2026 AI megarounds hit $240B — nearly triple 2024–2025 combined

by

Scatter chart of the 12 largest disclosed AI funding rounds from October 2024 to October 2026: 2026 megarounds dwarf earlier rounds, with OpenAI's $110B raise the largest and DeepSeek's $15B the only round from a Chinese lab

Twelve of the largest disclosed AI funding rounds since October 2024 total $321 billion — and $240 billion of that landed in 2026 to date, nearly triple the $81 billion raised across 2024 and 2025 combined. DeepSeek’s $15 billion raise, reported October 22 ahead of a planned Shanghai listing, is the largest ever by a Chinese AI lab, yet it ranks only fifth among 2026 rounds, behind OpenAI’s $110 billion, Anthropic’s $65 billion and $30 billion rounds, and xAI’s $20 billion. Frontier AI financing has moved from venture scale to infrastructure scale.

Sources & Notes

• 0 Comments

DeepSeek raised $15B in October at about $74B valuation after a $7.4B June round at about $52B

by

Scatter chart of DeepSeek's two 2026 funding rounds as unconnected discrete events: capital raised was $7.4B in the round closed 2026-06-16 and $15B in the round closed 2026-10-22, with reported valuations of about $52B and $74B respectively.

DeepSeek’s second private financing closed at about $15 billion with a reported valuation of roughly $74 billion going into the final close, after a June round that raised about $7.4 billion at a Caixin-reported post-money valuation of about $52 billion. A second, independent SI Plots dataset already showed DeepSeek’s reported private valuation rising sharply across 2026 toward about $74–75 billion (DeepSeek’s reported private valuation climbed about sevenfold in 2026); this chart adds the capital raised in each round and uses the Caixin post-money figure for June rather than other reported June marks. That paired raise-and-valuation path is the quantitative core of the SI News report that DeepSeek nearly doubled its fundraising target amid strong investor demand ahead of a possible Shanghai listing.

Sources & Notes

• 0 Comments

Meta’s quarterly capex rose nearly fivefold as free cash flow turned negative

by

Line scatter of Meta quarterly capital expenditures and free cash flow from 2023-03-31 to 2026-09-30 in USD billions; capex climbs from about 7 to 34.9 while free cash flow drops from about 12 to negative 2.1 in 2026.

Meta’s capital expenditures, including finance-lease principal, reached $34.9 billion in the third quarter of 2026, about 4.9 times the $7.1 billion it spent in the first quarter of 2023, while quarterly free cash flow fell from $12.4 billion in the first quarter of 2026 to negative $2.1 billion, its first negative quarter in this series. In other words, Meta’s AI data-center buildout now costs more than all the cash its operations generate, even though third-quarter revenue rose 26% year over year to $64.6 billion. That context matters for the SI News report that Meta raised the floor of its 2026 capex outlook to $135 billion and approved a new share class to help fund the expansion.

Sources & Notes

• 3 Comments

Anthropic’s annualized Q3 revenue of $67 billion matches the $65 billion run rate it claimed for July

by

Log-scale scatter plot of Anthropic's stated annualized revenue run rate from about $1 billion in January 2025 to $65 billion in July 2026, alongside booked revenue annualized from $4.6 billion for 2025 to about $67 billion in Q3 2026

Anthropic’s booked revenue has tracked the run-rate figures it publicized: it booked $4.6 billion in 2025 against a stated run rate of about $5 billion in August 2025, and $16.8 billion in Q3 2026 (about $67 billion annualized) against the $65 billion run rate reported for the end of July. That matters because a run rate is a company-chosen projection, and the booked quarters show the roughly 65-fold climb from about $1 billion in early 2025 was realized revenue, even as quarter-on-quarter growth slowed from 147% in Q2 to 45% in Q3. The same booked figures underpin the draft IPO prospectus covered in the SI News roundup, which pairs $16.3 billion of first-half revenue with an $80.7 billion net loss.

Sources & Notes

• 0 Comments

Independent test: Mercury 3 decodes about 3x faster than Claude Haiku 5.5 and 6x faster than GPT-6 Luna at similar code accuracy; Cerebras-served gpt-oss is 2.7x faster still

by

Scatter plot of median decode speed (tokens/s, log scale) versus code pass@1 for six model settings: Mercury 3 at 849-893 tok/s and 80.6-86.0%, Claude Haiku 5.5 at 324 tok/s and 86.1%, GPT-6 Luna at 132 tok/s and 83.2%, Gemini 3.5 Flash-Lite at 438 tok/s and 82.4%, gpt-oss-120b on Cerebras at 2,396 tok/s and 82.4%, with Inception's 1,068 tok/s claim as a dashed line

In an independent test run on 2026-10-20/21 over public APIs, Inception’s diffusion model Mercury 3 decoded long (~8k-token) answers at a median 849–893 visible tokens per second, 2.6–2.8x Claude Haiku 5.5 with thinking off (324 tok/s) and 6.4–6.8x GPT-6 Luna with reasoning off (132 tok/s), while scoring 80.6–86.0% pass@1 on 637 coding problems, level with those two models (86.1% and 83.2%) rather than ahead of them. The measured rate sits below Inception’s 1,068 tok/s launch figure (the tester rates the claim “close”, since client-side timing includes network overhead; OpenRouter separately lists a 247 tok/s median throughput for Mercury 3’s live traffic, a different all-traffic measure), and an ordinary autoregressive model, gpt-oss-120b served on Cerebras hardware, decoded 2,396 tok/s at 82.4%, so diffusion is one route to speed rather than the only one. That answers the “if the benchmarks hold up” caveat in today’s SI News item on Mercury 3: the speed lead over the big labs’ fast models is real but a constant factor, and code quality roughly matches them instead of beating them.

Sources: Bruce Li, “Mercury 3 vs. the Fast Autoregressive Models: A Technical Report” (2026-10-21), §5.1 table (pass@1, run r2-uc1-main-20261021T034053Z) and §5.4 table (long-answer visible decode speed, run r3-uc1-length-20261021T152238Z) · study code and configs (GitHub) · Inception, “Diffusion at Scale: Introducing Mercury 3” (1,068 tok/s claim) · OpenRouter, Mercury 3 model page (live P50 throughput, viewed 2026-10-21)
Original story: SI News: DeepSeek Open-Sources Its DeepInfer Inference Engine | US Bans DeepSeek From Civilian Agencies | Inception Launches Mercury 3 Diffusion Model
Notes: One point per model setting. Speed is the median visible decode rate (visible tokens after the first visible token, measured client-side from one US West host) on ~8,000-token answers, at each model’s lowest setting; Mercury 3 “low” includes hidden reasoning before its first visible token. Accuracy is pass@1 across 164 HumanEval+, 378 MBPP+ and 95 LiveCodeBench v6 problems, 5 samples each. Both runs come from the same single-author study; they compare products (model plus serving stack), not architectures. The x axis is logarithmic.

• 0 Comments

M3.1 Flash: 42.7 Intelligence Index at $0.22 per task

by

MiniMax’s M3.1 Flash (released 2026-10-19) scores 42.7 on the Artificial Analysis Intelligence Index at a measured $0.22 per index task. That puts it close to far pricier rivals — Qwen3.8 Max scores 45.4 at $5.41 per task, about 25 times the cost — in a plot of 20 frontier models from Artificial Analysis’s live leaderboard.

Source: SI News article

Data: Artificial Analysis

• 0 Comments

DeepSeek’s OpenRouter token share rose from 22% to 29% after MiniMax’s stealth Space Bunny left

by

Scatter chart of daily share of OpenRouter tokens, 2026-09-20 to 2026-10-19: MiniMax's stealth Space Bunny rises to 29.5% then leaves on 2026-10-05, after which DeepSeek's share rises from about 21% to 29-36%

OpenRouter’s daily token data show MiniMax’s anonymous Space Bunny Alpha climbing from 1.5% of all tokens on 2026-09-23 to 29.5% on 2026-10-04, 56.7 trillion tokens in total (matching OpenRouter’s own reveal figure), before it left the platform on 2026-10-05. DeepSeek’s combined share averaged 22.1% during Space Bunny’s run (2026-09-24 to 2026-10-04), jumped from 21.0% to 31.1% the day it left, and averaged 29.1% from 2026-10-06 to 2026-10-19; the timing fits users of the free stealth model moving to DeepSeek, though the aggregate data cannot show where individual users went. It matters for two stories in SI News’s MiniMax M3.1 Flash item and its report on the U.S. DeepSeek ban: the relaunched M3.1 Flash took 4.3% of tokens on its first (free) day, and DeepSeek, now barred from federal civilian agency systems, was still OpenRouter’s largest model developer at 31.5% of tokens on 2026-10-19.

Sources: OpenRouter rankings (daily tokens per model from the rankings-daily dataset, CC BY 4.0; Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-20T08:29:15Z), retrieved via the SF·SIX usage snapshot (SF·SIX most-used models page) · OpenRouter, Space Bunny Alpha model page (released 2026-09-23, revealed as MiniMax M3.1 Flash) · MiniMax, M3.1 Flash launch post · DHS order on DeepSeek, Federal Register 2026-21447
Original stories: SI News: Anthropic 5% shares into public benefit fund | MiniMax model M3.1 Flash with 1M context | Hybrid Reasoning Agents · SI News: U.S. bans DeepSeek from federal civilian agency systems
Notes: Shares are each developer’s tokens divided by all OpenRouter tokens that day (top-50 models plus the aggregated long-tail row), UTC days. DeepSeek includes all of its listed models. Token counts come from each provider’s own tokenizer, so cross-developer shares are approximate, and the data cover OpenRouter traffic only, not global use. Space Bunny Alpha and the 2026-10-19 M3.1 Flash listing were free to use.

• 0 Comments

AI agents’ 80%-reliable task length runs about 5x shorter than their 50% horizon

by

Scatter plot of METR 50% and 80% success time horizons in human-expert hours by model release date, 2024-05 to 2026-04

Across 19 models released between 2024-05-13 and 2026-04-07 in METR’s Time Horizon 1.1 data, the task length a model completes with 80% predicted success is a median 5.3x shorter (range 4.0x to 10.3x) than the length it completes with 50% success. The most recent point, an early Claude Mythos Preview, reaches 17.4 human-expert hours at 50% but only 3.1 hours at 80%, and METR cautions that estimates above 16 hours are unreliable. The gap matters because attackers can tolerate frequent failure: in the AI-agent hacking campaign covered in today’s SI News brief, Reuters reports that the agents attempted more than 1,200 sites and compromised more than 100, while businesses deploying agents generally need the reliable end of the curve.

Sources: METR, Task-Completion Time Horizons of Frontier AI Models (Time Horizon 1.1 data file; p50 and p80 horizon estimates and release dates) · Reuters (republished by RTÉ) · NorForge, Inside Cairn
Original story: SI News: Agents that hack, Anthropic’s public-benefit fund, Disney’s Character.AI hire
Note: METR’s horizon measures how long the tasks take human experts, not how long an agent runs. Points are point estimates; METR’s confidence intervals are wide (Claude Mythos Preview at 50%: 8.5 to 55 hours).

• 0 Comments