Answer card: Reflection AI announced Beam, a 501B parameter open-weight mixture of experts model, with weights promised later in October 2026.

Reflection Beam is a 501B open model with no weights yet

6 October 2026
Answer card: Aleph Alpha released Kolibri-1 on 3 October 2026 as an Apache 2.0 mixture of experts model with 78.1 billion total and 3.46 billion active parameters, served at 1 million tokens of context but trained on sequences of up to 256 thousand tokens.

Kolibri-1 serves 1M tokens but was trained to 256K

4 October 2026
Answer card on Gemini 4 Argon pricing and gated access, announced by Google on 30 September 2026.

Gemini 4 Argon costs $2 and $10 now, $4 and $20 later

2 October 2026
Answer card for GPT-6.1 Sol showing 2 dollars in and 10 dollars out per million tokens, unchanged from GPT-6 Sol, with cached input at 10 cents.

GPT-6.1 Sol costs what GPT-6 Sol did, and only the cache got cheaper

30 September 2026
Answer card for Claude Sonnet 5.5 showing 2 dollars in and 10 dollars out per million tokens, unchanged from Sonnet 5, and between_tools replacing disabled thinking.

Claude Sonnet 5.5 keeps $2 and $10 and retires thinking disabled

29 September 2026
Answer card: Xiaomi retrained MiMo-V2.6 to stop repeating tool calls, API swapped on 25 September 2026, MIT weights on 27 September, same model names.

MiMo-V2.6-Pro was quietly retrained to stop looping on tool calls

28 September 2026
Answer card: Anthropic committed $11.6 billion over seven years to Akamai Cloud for CPU workloads only, with revenue from the second half of 2027.

Anthropic’s $11.6B Akamai deal buys CPUs, not GPUs

27 September 2026
Answer card: Google Suncatcher MVP satellite, four Trillium TPUs on about one kilowatt, launching on SpaceX Transporter-18, reported for 1 October 2026.

Google’s first Suncatcher satellite flies four TPUs on 1 kW

25 September 2026
Answer card for Claude Opus 5.5: 4 dollars in and 20 dollars out per million tokens, down from 5 and 25, with cache reads at 20 cents.

Claude Opus 5.5 drops to $4 and $20, and breaks four Opus 5 habits

23 September 2026
The official xAI announcement card for Grok 4.7, white type on a dark grey and navy gradient.

Grok 4.7 keeps $2 and $6, and its gains over 4.6 are xhigh versus high

22 September 2026
Answer card stating that Qwen-Image-2.1, released on 20 September 2026, ships open weights with a 7 billion parameter diffusion transformer, a Qwen3-VL 8B text encoder and an RGBA VAE totalling about 33 gigabytes in BF16, under the Qwen Research License that limits use to research or evaluation and requires a separate commercial licence, unlike the Apache 2.0 licence of Qwen-Image 1.0.

Qwen-Image-2.1 brings the weights back, but not the Apache licence

21 September 2026
Answer card stating that Ternary Bonsai 2 27B, released by PrismML on 17 September 2026 under Apache 2.0, packs Qwen3.8 27B into 5.95 gigabytes at 1.72 bits per weight, keeps 98.2 percent of the 14-benchmark average, about 75 percent on SWE-bench Verified and Terminal-Bench 2.1, and needs PrismML's llama.cpp fork to run.

Does Bonsai 2 27B really keep 98% of Qwen3.8 in 5.95 GB?

20 September 2026
Wednesday, October 7, 2026
No Result
View All Result
No Result
View All Result

Tag: mixture of experts

  • Trending
  • Comments
  • Latest

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result

Copyright © 2026 Stephane Cardon.