Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-09-26 · source-backed
Mica is Qwen3.5-4B with a rank-16 LoRA on all 32 layers, trained on 77,732 rows on rented RTX 3090s. It speaks TypeSafe's /v1/systemone format, so Jev clients work against it. On the author's held-out set it scores 67.0 against Jev 1.13's 74.1 and Kev 4B's 57.0, dropping to 64.9 on JevBench hard under the official runner. In the Minecraft demo it scored candidate commands at 90-150ms each on a 3090. Commenters correctly noted that reading label logits is still greedy decoding of one token, which is also what makes it cheap.
Each link below shares sources, entities, or timing with this story.
Benchmark Heaven's JevBench scores 534 fixed decisions on four equally weighted axes: intelligence above chance, calibration, speed and cost. Jev 1.13.0 leads at 74.4 for $0.040, SemIf (Qwen3.5-4B) follows at 73.1 for about $0.022, diffusion-Gemma djev at 73.0. GPT-5.6 Luna on...
Google DeepMind released Gemma 4 on April 2 with four model sizes (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0. Multimodal (text, vision, audio). 256K context. Native thinking and tool-calling optimized for agentic workflows. Day-zero ecosystem support across vLLM, llama.cp...
The models are good. The license is the real story. Google released Gemma 4 on April 2 with four variants: E2B, E4B, 26B MoE, and 31B Dense. All built on the Gemini 3 architecture. The 31B Dense variant claimed #3 on Arena AI's text leaderboard, beating models 20x its size. Th...
PortLLM claimed training-free, data-free transfer of LoRA patches onto updated base models, but only over short horizons and without theoretical grounding. This study runs 10 continual-pretraining steps on Mistral, Gemma, and Qwen and finds portability persists long-run, meani...
Jared Palmer published Kev September 17 on Qwen3.5, Apache-2.0 with training code and frozen eval suites. One request carries yes/no, multiple-choice and rating questions that share input text but can't read each other, and the API matches TypeSafe's System One so their Python...
Alibaba released Qwen3.6-27B on April 22. Dense architecture. Open weights. 77.2% on SWE-bench Verified, within 3.7 points of Claude Opus 4.6. On SkillsBench, it scores 48.2% versus its own 397B MoE predecessor's 30.0%. That's a 77% improvement with 14.8x fewer parameters. Let...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.