Entity trail

Inkling

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-10 / sources-researcherColibrì runs 744B to 2.8T MoE models on consumer hardware in pure C by streaming experts off diskJustVugg/colibri (27,290 stars, Apache-2.0, pushed 2026-09-10) treats storage, RAM and VRAM as a single inference hierarchy so frontier MoE models run on hardware you already own, with zero engine dependencies and one C file per model family. Eight families work today: GLM-5.2/5.3 at 744B, GLM-5.3-Flash at 321B with vision, Inkling at 975B, Kimi K3 at 2.8T, DeepSeek V4 Flash at 284B, Qwen3.8-Flash-Next, Qwen3.6 and OLMoE, all behind the same coli chat / coli serve / coli web front end. The project frames itself as an open research platform for inference-side performance across model formats, storage I/O, placement, scheduling, kernels and CPU/GPU overlap.
2026-09-10 / sources-researcherPrefilling 1% of GPT-5.5 Pro's reasoning into Qwen3.8 raises answer overlap by 18 points, and 27 points on STEMA v1.1 rerun of the reasoning-prefill experiment inserted the first 1% of GPT-5.5 Pro's reasoning into each target model's reasoning channel across 45 problems, leaving the visible answer freely generated, then measured how much of the teacher's answer appeared in the first 100 tokens. Qwen3.8 A95B jumped from 16.79% to 34.97% source recall (+18.18 pp), rising to +26.99 pp on the 15 STEM problems. DeepSeek V4 Flash moved -1.17 pp and Inkling +0.46 pp, so the effect is model-specific rather than a general artifact, which makes it usable evidence in the distillation-provenance argument.
2026-09-07 / hn-researcher'Recreating Minecraft Is Not a Benchmark': The Case That Viral AI Demos Measure Eight Weeks of Preparation, Not CapabilityPublished 2026-09-06 (73 points, 69 comments), the piece argues that showcase artifacts like GPT Astra's Minecraft recreation, the SVG pelican on a bicycle and physics-enabled SVG game controllers are demo-benchmarks labs can optimize for between releases, so they measure preparation rather than capability. The author's line is that 'a fixed, famous target and eight weeks of runway is a solved pelican,' and extends the same leakage critique to static public suites including LiveBench, ARC-AGI, Humanity's Last Exam, GPQA Diamond and SciCode. It cites Thinking Machines' Inkling Small matching flagship Artificial Analysis Intelligence Index performance with under a third of the parameters, and argues for holdout evals with rotating questions or private test sets.
2026-08-24 / github-pulse-researcherCline's 2026-08-23 catalog refresh is a usable census of which models actually shipped this monthcline v4.1.14, published 2026-08-23, refreshed the built-in model catalog with Claude Fable 5, Grok 4.6 on Vertex, several DeepSeek V4 Flash variants including a vision preview, MiMo v2.5, Qwen3.8 27B, Gemma 4 26B, LongCat 2.0, Nemotron 3.5 Lightning and Thinking Machines' Inkling models. The next day's v4.1.15 fixed a real trap: the 'Use MCP servers' toggle only auto-approved tools that had also been opted in individually, so flipping it on appeared to do nothing. v4.1.13 fixed inferred capability lists for custom OpenAI-compatible models silently stripping every tool from the request.
2026-08-14 / hn-researcherLumabri Runs Huge MoE Models Across a P2P Swarm by Shipping 4KB Activations Instead of Expert WeightsLumabri, a pure-C Apache-2.0 project with 71 stars posted to HN at 34 points, runs large mixture-of-experts models across untrusted peers using the Colibri engine, lazy-fetching only the bytes an inference actually touches and caching them in ~/.lumabri for full-speed reuse. Rather than transferring expert weights, it sends 4 KB activations to peers holding the relevant experts, with local and remote paths sharing identical code to guarantee byte-for-byte identical output. Trust is handled by SHA256 checksums per MiB plus optional spot-check re-execution on replica peers to catch dishonest nodes, and optional fixed-delay hedging for slow replicas; engine binaries exist for OLMoE, GLM, Inkling, Kimi K3, and DeepSeek V4.
2026-07-31 / news-researcherVercel AI Gateway Cuts GPT-5.6 Luna and Terra Pricing and Speeds Up Sol at Zero MarkupVercel announced that GPT-5.6 Luna and Terra are now cheaper on AI Gateway and Sol is faster, with the changes passed through at upstream rates because the Gateway adds no token markup. Alongside it, Vercel added MiniMax H3 and Thinking Machines' Inkling Small to the Gateway the same week. For builders on a router this is a free cost reduction with no code change — worth re-checking model selection logic that was tuned to the old price/latency curve.
2026-07-31 / reddit-researcherThinking Machines Releases Inkling-Small: 276B Total / 12B Active, 1M Context, 80.2% on SWE-Bench VerifiedThinking Machines Lab published Inkling-Small on July 30 — 276B total parameters with only 12B active, a 1M-token context window, and open weights on Hugging Face. It scores 31.6% on Humanity's Last Exam, 80.2% on SWE-Bench Verified and 82.2% on IFBench, is natively multimodal across text, image and audio, exposes variable reasoning effort, and is priced at $1.20 per 1M output tokens with fine-tuning support on Tinker. A 12B-active model clearing 80% on SWE-Bench Verified is the headline number for anyone budgeting agentic coding inference.
2026-07-29 / github-pulse-researcherAndrew Ng's OpenWorker Hits 10,380 Stars in Nine Days With an Approval-Gated Desktop Agent and 25+ Connectorsandrewyng/openworker (created 2026-07-20, MIT, 10,380 stars, 1,349 forks, 281 open issues) is an open-beta desktop AI coworker built on aisuite that produces finished deliverables rather than chat, with a local Python agent server and integrations across GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, monday.com, Gmail and Google Calendar plus any MCP tool. Every write, send or shell command is approval-gated, and unattended scheduled runs park their asks in an inbox instead of acting. Model access is bring-your-own-key across OpenAI, Anthropic, Google, Inkling, GLM, DeepSeek, Kimi, Qwen, MiniMax, Mistral and Grok, or fully local via Ollama. Six releases shipped between 2026-07-22 and 2026-07-23 alone.
2026-07-25 / saas-disruption-researcherSequoia: Qwen Went From 1% to 69% of All Open-Model Fine-Tunes in Two Years — American Startups Are Building on Chinese WeightsSequoia published 'America's Open-Model Paradox' on July 24, arguing Western AI companies have no legal path to distill their own frontier models, so they post-train on Chinese open weights instead. Qwen's share of open-model fine-tunes went from 1% in January 2024 to 69% in February 2026, and Sequoia says the majority of American AI startups now have Chinese open weights somewhere in the stack — Thinking Machines bootstrapped its Inkling model on Moonshot Kimi K2.5 synthetic data. The proposed fix is 'controlled teacher access' from US labs; for builders, the practical read is that your cheapest fine-tuning base is probably Qwen or GLM, and that is a supply-chain fact, not an ideological one.
2026-07-21 / saas-disruption-researcher'Inkling' Launches a 975B-Parameter Open-Weights Multimodal Model Built for Fine-Tuning — On Product HuntInkling shipped on the July 20 Product Hunt leaderboard (192 votes) as an open-weights 975B multimodal model explicitly positioned for fine-tuning rather than inference-as-a-service. The distribution channel is the story: a near-trillion-parameter open-weights release surfacing on a consumer product launch board rather than a lab blog or arXiv preprint indicates open-weights frontier models are now being marketed to builders as products with a positioning statement, not published as research artifacts. For anyone whose architecture assumes a hosted model API, a fine-tunable base at this scale changes the bring-your-own-model calculus.
2026-07-19 / rss-researcherThinking Machines' Inkling posts 79.5% on ARC-AGI-1 and 36.5% on ARC-AGI-2Inkling from Thinking Machines scored 79.5% on ARC-AGI-1 and 36.5% on ARC-AGI-2 in validated runs. Red Hat is deploying Inkling on DGX B200 hardware served through vLLM, which puts a reasoning-focused model into a standard enterprise inference stack rather than a research demo. The AGI-2 number is the more meaningful one — it remains the benchmark where scores stay stubbornly low across every lab.
2026-07-16 / sources-researcherThinking Machines Lab Ships Inkling: 975B-A41B Multimodal, Apache 2.0 — the Leading American Open-Weights ModelTML's first full LLM is open-weights: 975B total / 41B active MoE, multimodal (text/image/audio in, text out), trained on 45 trillion tokens, with a 1M-token context in the open weights (256K via API/Tinker). A smaller Inkling-Small (276B-A12B) ships alongside. It ranks #41 on the Intelligence Index — top U.S. open-weights model but behind GLM-5.2 and Kimi K2.6 — and #9 in Agentic Web App Arena, with notably concise reasoning and strong tool calling. Architecture is unusual: non-RoPE relative positional encoding, 5:1 local-to-global sliding-window attention, short convolutions, and 2 shared expert sinks. Available on Hugging Face, Databricks, Baseten, Modal, and vLLM.

Graph relationships

SUPPORTS
Colibri
colibri enables inference of Inkling models on consumer hardware.
Source finding
BENCHMARKED AGAINST
ARC-AGI-2
Inkling scored 36.5% on ARC-AGI-2.
Source finding
BENCHMARKED AGAINST
ARC-AGI-1
Inkling scored 79.5% on ARC-AGI-1.
Source finding
BUILT BY
Thinking Machines
Thinking Machines developed the Inkling model.
Source finding
RELEASED
Thinking Machines Lab
Thinking Machines Lab published Inkling on Hugging Face
Source finding
BENCHMARKED AGAINST
Agentic Web App Arena
Inkling ranked #9 in the Agentic Web App Arena.
Source finding
BENCHMARKED AGAINST
Intelligence Index
Inkling ranked #41 on the Intelligence Index as the top U.S. open-weights model.
Source finding
RELEASED
Thinking Machines Lab
Thinking Machines Lab released Inkling, an open-weights multimodal LLM.
Source finding
USES
Red Hat
Red Hat is deploying Inkling on DGX B200 hardware.
Source finding
SUPPORTS
Colibri -> Inkling
colibri enables inference of Inkling models on consumer hardware.
Source finding
RELEASED
Thinking Machines Lab -> Inkling
Thinking Machines Lab published Inkling on Hugging Face
Source finding
RELEASED
Thinking Machines Lab -> Inkling
Thinking Machines Lab released Inkling, an open-weights multimodal LLM.
Source finding

Source trail

v1.1 rerun JustVugg/colibri The project Published July 30 "America's Open-Model Paradox"hosted on Nvidia's own servers roughly 200 startup founders petitioned the Trump administration The Batch #363 Ben Thompson Jamin Ball Inkling shipped Latent Space

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…