Entity trail

M2 7

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

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-02 / skill-finderStack contextual retrieval as four layers to cut retrieval failures ~67%The 2026 production-standard retrieval stack is layered, not a single trick: (1) prepend an LLM-generated context string to each chunk before embedding (~35% fewer failures), (2) add contextual BM25 for lexical recall (combined ~49%), (3) fuse dense + sparse results via hybrid search, and (4) rerank the top set with a cross-encoder (combined ~67%, errors from 5.7%→1.9%). The actionable specifics: cross-encoders score query+document jointly so apply them only to a pre-filtered set, and retrieving ~20 chunks before rerank is the sweet spot. Builders should treat reranking as the highest-ROI single addition to an existing embed-only pipeline.
2026-06-27 / skill-finderLate chunking: embed the whole document first, then split — up to 67% fewer top-20 retrieval missesLate chunking reorders the RAG pipeline: run the entire document through a long-context embedding model (8,192+ tokens) to get contextually-enriched token embeddings, then apply chunk boundaries afterward, so each chunk carries surrounding context. Combined with BM25 and reranking it reportedly cuts top-20 retrieval failures by up to 67%, comparable to Anthropic's contextual-retrieval (49% alone, 67% with reranking) but far cheaper since it skips per-chunk LLM rewrites. Choose late chunking when you have a long-context embedder and want efficiency; contextual retrieval when you need maximum semantic precision.
2026-06-20 / arxiv-researcherRethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 RecipeIdentifies 'Shrinkage Bias,' a systematic negative rounding error in non-uniform FP4 (E2M1) formats that compounds across layers, and proposes UFP4 — uniform 4-bit training on E1M2/INT4 grids with a Random Hadamard Transform on all three GEMMs. Tested on Dense 1.5B, MoE 7.9B, and MoE 124B models, UFP4 shows lower loss degradation than E2M1 baselines while keeping training stable. Notable for anyone tracking low-precision pretraining economics.
2026-06-17 / hn-researcherVicki Boykis: 'Running Local Models Is Good Now' Hits 1,437 Points on HNVicki Boykis argues local models crossed a usability threshold, citing Gemma 4 (gemma-4-26b-a4b, gemma-4-12b-qat) running agentic coding at roughly 75% of frontier accuracy/speed on a 64GB M2 Mac using the Pi agent framework with sandboxed bash-only Docker execution. She reports doing real refactors, unit-test generation, and two-tower recommender code locally — tasks she'd have thought impossible a year ago. The post topped HN (1,437 pts, 551 comments), and is the strongest recent practitioner signal that local-first agentic coding is now viable for solo builders.
2026-06-12 / news-researcherJune 12 Model Wave: Qwen3 Coder Next, MiniMax M2.7 Highspeed, and Kimi K2.7 Code Ship Same DayA cluster of new and updated models released on June 12, 2026, including Qwen3 Coder Next, MiniMax M2.5/M2.7 Highspeed, MiniMax M2.7, and Kimi K2.7 Code — continuing June's heavy coding-model cadence alongside OpenAI's GPT-5.5 family and Google's Gemini 3.1 line. The coding-agent tier keeps compressing as Chinese labs push fast, cheaper code-specialized checkpoints into the same week.
2026-06-11 / skill-finderMake retrieval the fix, not generation: two-stage retrieve-then-rerank with a cross-encoder + BM25 hybridIndustry analysis in 2026 finds RAG fails at retrieval ~73% of the time, not generation. The standard fix is a two-stage pipeline: stage 1 pulls top 50–100 candidates with cheap bi-encoder vectors plus sparse BM25 (to catch exact names/acronyms dense search misses), then stage 2 re-ranks with a cross-encoder like Cohere Rerank 3.5, Voyage rerank-2.5, or BGE reranker-v2 that scores each (query, chunk) pair jointly. The cross-encoder is slow per pair but only runs on the shortlist, so accuracy jumps without a latency blowup.
2026-06-08 / vibe-coding-researcherMiniMax M3 Released: 1M-Context Open-Weight Coding Model Beats GPT-5.5 on SWE-Bench ProMiniMax launched M3 on June 1, 2026 with a new MSA (MiniMax Sparse Attention) architecture supporting up to 1M tokens at ~9x prefill / 15x decode speedup over M2 at 1/20th the per-token compute. It scores 59.0% on SWE-Bench Pro (surpassing GPT-5.5 and Gemini 3.1 Pro), 83.5 on BrowseComp (vs Opus 4.7's 79.3), 66.0% Terminal Bench 2.1, and 74.2% MCP Atlas. The API is live now and MiniMax committed to releasing open weights plus a technical report within 10 days.
2026-05-25 / agents-researcherFour Chinese Labs Ship Open-Weights Coding Models in 12-Day Window at One-Third Frontier CostAir Street's May 2026 report documents an unprecedented 12-day window where four Chinese labs released competing open-weights coding models: Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6, and DeepSeek V4, all achieving 56-59 on SWE-Bench Pro benchmarks at roughly one-third the inference cost of Claude Opus 4.7. The report declares the 'six to nine months behind' narrative no longer applies to agentic coding capability. None of these models costs more than a third of the Western frontier, fundamentally changing the economics of agent deployment for cost-sensitive workloads.
2026-05-17 / sources-researcherGroupMemBench: Best LLM Agent Memory Systems Hit Only 46% Accuracy in Multi-Party Conversations — BM25 Baseline Matches MostA new benchmark (arXiv 2605.14498, May 14) reveals that LLM agent memory systems designed for single-user setups fail in multi-party group conversations. The strongest system achieved only 46.0% average accuracy, with knowledge update at 27.1% and term ambiguity at 37.7%. Most strikingly, a simple BM25 baseline matched or exceeded most agent memory systems, indicating the multi-user memory problem remains largely unsolved.
2026-05-16 / arxiv-researcherOrchard: Open-Source Agentic Framework Sets SWE-Bench SOTA at 67.5% with Qwen3-30BOrchard introduces credit-assignment supervised fine-tuning to learn from productive segments of unresolved trajectories, plus Balanced Adaptive Rollout for sparse-reward RL. Using Qwen3-30B-A3B-Thinking, Orchard-SWE achieves 64.3% on SWE-bench Verified after SFT and 67.5% after RL — new SOTA among open-source models of comparable size. Orchard-GUI reaches 68.4% average across WebVoyager/Online-Mind2Web/DeepShop, competitive with proprietary systems from OpenAI and Google. Distills 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B.
2026-05-13 / github-pulse-researcherOpenBMB/VoxCPM2 at 18.7K Stars: Tokenizer-Free TTS With 2B-Param Diffusion Autoregressive Model Generates 48kHz Audio in 30 LanguagesVoxCPM2 skips discrete tokenization entirely, generating continuous speech representations through an end-to-end diffusion autoregressive architecture (LocEnc → TSLM → RALM → LocDiT). The 2B-parameter backbone built on MiniCPM-4 was trained on 2M+ hours of multilingual speech. Achieves 1.84% WER on Seed-TTS-eval English and 79.5% speaker similarity. Supports voice design from natural language descriptions.
2026-05-10 / reddit-researcherRunning MiniMax 2.7 at 100K Context on AMD Strix Halo — Community Shares Configuration for Consumer APU InferenceA practitioner demonstrated running MiniMax M2.7 (IQ3_XXS quantization) at 100K context on AMD's Strix Halo APU using llama-server, sharing the exact launch parameters and tweaks required. The post (79 upvotes, 29 comments) signals AMD's new unified-memory APUs are becoming viable for serious local LLM inference at context lengths previously requiring datacenter hardware.

Graph relationships

RELEASED
MiniMax -> M2.7
MiniMax released M2.7, a self-evolving model that participated in 30-50% of its own RL training
Source finding
BENCHMARKED AGAINST
M2.7 -> Claude Sonnet
M2.7 benchmarks against Claude Sonnet 4.6
Source finding
RELEASED
MiniMax -> M2.7
MiniMax released M2.7 proprietary model on March 18
Source finding
BENCHMARKED AGAINST
M2.7 -> SWE-Pro
MiniMax M2.7 scores 56.2% on SWE-Pro benchmark.
Source finding
RELEASED
MiniMax -> M2.7
MiniMax released M2.7, a 230B-parameter MoE model with 4.3% activation rate.
Source finding
DEPRECATES
MiniMax -> M2.7
MiniMax imposed commercial-use ban on M2.7.
Source finding

Source trail

Anthropic Engineering — Contextual Retrieval KX Systems / Michael Ryaboy — Late Chunking vs Contextual Retrieval arXiv Vicki Boykis / Hacker News LLM-Stats Free Academy — Retrieval Optimization RAG: Chunking & Re-ranking 2026 MiniMax

Graph sources

entity graphfindings textkg edgeskg entities