This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

On Nov 5 (US time), Mira Murati's Thinking Machines Lab released Inkling 2 Small. The official page is clear: 594B total / 17B activated sparse MoE, image input, adjustable reasoning. The biggest release in this window, and a lab this column hasn't covered before.
What's worth remembering isn't the score but a new open-source flow - the official wording: serve it closed-weight on Tinker first, run external safety tests and red-teaming, then "next week" release full Apache 2.0 weights with checkpoints. "Open source" is split into two steps: let you use it, then let you take it.
Vendor-claimed 86.3% on hard coding problems (unreplicated). The takeaway is the process: a "closed first, open later" cadence that lets a lab ship fast while still keeping the eventual open-source promise.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| Kanana-2-v-decide-3b (Kakao, open source) | 11-6 | Decision-specialist | ~3.67B (0.49B vision encoder + 3.18B language tower). It never generates text: encodes question and candidates into 2,560-dim vectors, cosine similarity + softmax outputs choice/noul/score with probabilities, candidate vectors cacheable across calls. Vendor: Korean Decision Bench 72.50 (15,909 questions), Decision Index 61.34; one H200 averages 0.04s, claimed ~43x GPT-6 Luna. Kanana Open License. Note: only judges given candidates, no video, probabilities relative to the candidate set. |
| Cola2 (ByteDance Seed, open source) | 11-6 | Continuous-latent foundation model | ~9B, not autoregressive: a dynamic-compression VAE chunks text into semantic segments compressed to continuous latents, a ~4B block-causal diffusion Transformer generates latent blocks via flow matching, a span-causal decoder turns them back into words. 346.03B tokens; self-reported 8-task mean 60.10, clearly below Llama-3-8B's 66.25 and Qwen3-8B-Base's 75.87. Apache 2.0 code, separate weight terms, diffusion part self-admittedly unconverged - a direction signal, not production-ready. |
| PrimeSim (PrimebotTech, open source) | 11-6 | Action-conditioned world model | Code plus two Apache-2.0 checkpoints (~10.1GB each), paper arXiv 2611.04870, base Wan2.2-TI2V-5B: first frame per camera + language + robot action trajectory yields synced next frames across three views. Trained on 22 robot platforms, 30,110 hours, a unified 31-dim cross-embodiment action representation. ~10Hz on one A100. Not a single-file package - you still fetch Wan2.2's text encoder/VAE/tokenizer; prepare A100/H100 80GB. |
| Cosmos 3 four checkpoints (NVIDIA, open source) | 11-5 | World model / simulation | Four post-trained Cosmos 3 Nano checkpoints, 16B trainable params each, OpenMDW1.1 commercial license: Transfer-Auto (offline driving data, up to 11 cameras + LiDAR, 400 frames); Sim-Auto (streaming distillation for closed-loop sim); Sim-Transfer (physics-engine layouts rendered photoreal); Sim-Bimanual (next-frame from bimanual actions, 48.1 fps / 166ms on one GB300). Nano generation needs 58GiB BF16 / 44GiB FP8. |
| Kimi K3.1 Pro weights (Moonshot, re-verified) | 11-6 | Open weights live | Independent re-check: no gate, 112 files / 96 safetensors shards, ~11.7K downloads / 944 likes; 2.8T total, 104B active, 93 layers, 16 of 896 experts + 2 shared, 1,048,576 context, MXFP4 weights / MXFP8 activations. Two hard license lines: MaaS with >$20M annual revenue (including affiliates, trailing 12 months) needs a separate Moonshot contract; products with >100M MAU or >$20M monthly revenue must carry the Kimi brand. |
"Closed-first, open-next-week" is becoming the standard release cadence for frontier labs.
(Data from official announcements and public reports, 2026-11-07; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa