[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fd4zlg463p4jo":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":24,"bodyZh":28,"bodyEn":29,"sourceName":30,"sourceUrl":31,"updatedAt":32},"recursive-flint-40b-2026-10-22","NEWS","2026-10-22T10:00:00.000Z","Recursive 发布 Flint：39.8B-A2.5B 编码 MoE，SWE-bench Pro v2 拿下 89.7%","Recursive ships Flint: a 39.8B-A2.5B coding MoE hitting 89.7% on SWE-bench Pro v2","10 月 21 日 Recursive 开源 Flint（Apache-2.0）——39.8B 总参、每 token 约 2.5B 激活的编码 agent 专用 MoE。发布评测：Terminal-Bench 2.1 54.9%、SWE-bench Pro v2 89.7%、SWE-bench Verified 74.2%。49 层里 35 层滑动窗口（513 token），社区 GGUF Q4_K_XL 25.1GB。TechCrunch\u002FWIRED 报道称主体由 AI 在不到 4 个月内构建。站内已收录并新增 22 条实测。","Recursive open-sourced Flint on Oct 21 (Apache-2.0), a coding-agent MoE with 39.8B total parameters and about 2.5B active per token. Release evaluations: Terminal-Bench 2.1 54.9%, SWE-bench Pro v2 89.7%, SWE-bench Verified 74.2%. Of its 49 layers, 35 use a 513-token sliding window; community GGUFs run Q4_K_XL at 25.1GB. TechCrunch and WIRED report it was largely built by AI in under four months. Logged on-site with 22 new runs.",[12,15,18,21],{"id":13,"name":14},6,"Qwen3.6-27B",{"id":16,"name":17},47,"Qwen3.6-35B-A3B",{"id":19,"name":20},92,"Flint-40B-A2.5B",{"id":22,"name":23},94,"mammoth-1",[25],{"id":26,"name":27},13,"NVIDIA RTX 5090","# Recursive 发布 Flint：39.8B-A2.5B 编码 MoE，SWE-bench Pro v2 拿下 89.7%\n\n10 月 21 日，推理基础设施公司 Recursive 开源了 **Flint**（Apache-2.0）——一个专为编码 agent 设计的 MoE：39.8B 总参，每 token 只激活约 2.5B（含输出头；纯 body 约 2.2B，官方记号 40B-A2.5B）。同日 TechCrunch 与 WIRED 的报道给出一个更抓眼球的注脚：这个模型的主体由 AI 在不到 4 个月内构建完成。\n\n## 发布评测（Recursive 自测口径）\n\n| 基准 | 得分 |\n|---|---|\n| Terminal-Bench 2.1 | **54.9%** |\n| SWE-bench Pro v2 | **89.7%** |\n| SWE-bench Verified | 74.2% |\n\n口径注意：温度 1.0 \u002F top-p 0.95，agentic 分数为 ≥5 次 pass@1 均值；Terminal-Bench 用 Terminus 2、SWE-bench 用 mini-swe-agent。Recursive 明确说明其 Terminal-Bench 口径把「靠网页检索抄到解法」计为失败——各家开发者自报分数未必做了同样的修正，横向比较要留个心眼。\n\n## 架构：滑窗为主的 49 层\n\n- 49 层 = 2 层稠密 + 47 层 MoE；256 路由专家选 6 + 常驻共享专家\n- 注意力 35 层滑动窗口（513 token：当前 + 前 512）+ 14 层全注意力；全注意力层不加位置编码，滑窗层用 RoPE\n- 每头 QK RMSNorm + attention sink——长上下文稳定性的常规组合\n- 上下文 262,144 token，词表 150,144，训练截止 2026-06\n- BF16 权重约 80GB，官方提供 B300\u002FH200 的 SGLang 服务套件\n\n滑窗为主的设计是「编码 agent」定位的直接产物：代码库检索 \u002F 修复循环里大量前缀可复用，513 token 窗口把 KV 压到极小，注意力开销让位给专家带宽。\n\n## 社区 GGUF（unsloth，10-21 已铺开）\n\n| 量化 | 文件大小 |\n|---|---|\n| UD-IQ1_M | 11.4 GB |\n| UD-Q2_K_XL | 13.6 GB |\n| UD-Q3_K_XL | 18.5 GB |\n| **UD-Q4_K_XL** | **25.1 GB** |\n| UD-Q5_K_XL | 29.9 GB |\n| UD-Q6_K_XL | 36.3 GB |\n| Q8_0 | 42.3 GB |\n\nQ4_K_XL 25.1GB 意味着单张 32GB 的 5090 就装得下 40B 级编码模型——物理上成立；不过站内显存口径是保守加 padding 的公式（含 KV\u002F运行时余量），本波实测主要落在 ≥60GB 级硬件上。\n\n## 站内收录\n\n**Flint-40B-A2.5B**（id 92）已入库，新增 **22 条 L2 实测**（llama.cpp \u002F vLLM \u002F Ollama \u002F Strata），全池均值 156.2 tok\u002Fs：\n\n| 硬件 | 量化 | 实测 |\n|---|---|---|\n| RTX PRO 6000 96GB | Q3_K_S | 343.5 tok\u002Fs |\n| H100 | IQ2_M | 326.0 tok\u002Fs |\n| M5 Ultra | IQ1_S \u002F Q2_K | 324.8 \u002F 230.7 tok\u002Fs |\n| RTX Spark（116 GiB Linux 池） | Q4_K_M | 70-76 tok\u002Fs |\n| Strix Halo 128GB | Q4_K_S | 63.8 tok\u002Fs |\n\nA2.5B 激活让它成为本批最快的模型之一——MoE 解码在带宽账上只读激活专家，2.5B 激活比同门 40B 稠密快一个数量级。","# Recursive ships Flint: a 39.8B-A2.5B coding MoE hitting 89.7% on SWE-bench Pro v2\n\nOn Oct 21, inference-infrastructure company Recursive open-sourced **Flint** (Apache-2.0) — a MoE built for coding agents: 39.8B total parameters with roughly 2.5B active per token (including the output head; the body alone is ~2.2B — the official designation is 40B-A2.5B). TechCrunch and WIRED added a grabbing footnote the same day: the model was largely built by AI in under four months.\n\n## Release evaluations (Recursive-measured)\n\n| Benchmark | Score |\n|---|---|\n| Terminal-Bench 2.1 | **54.9%** |\n| SWE-bench Pro v2 | **89.7%** |\n| SWE-bench Verified | 74.2% |\n\nMethodology notes: temperature 1.0 \u002F top-p 0.95, agentic scores are average pass@1 over at least five attempts; Terminal-Bench uses Terminus 2 and SWE-bench uses mini-swe-agent. Recursive explicitly counts solutions obtained through web retrieval as failures on Terminal-Bench — developer-reported comparison scores may not apply the same adjustment, so treat cross-source comparisons with care.\n\n## Architecture: a 49-layer, sliding-window-first design\n\n- 49 layers = 2 dense + 47 MoE; 256 routed experts with 6 selected per token plus an always-active shared expert\n- Attention is 35 sliding-window layers (513 tokens: the current one plus the previous 512) plus 14 full-attention layers; full-attention layers carry no positional encoding, sliding-window layers use RoPE\n- Per-head QK RMSNorm plus attention sinks — the usual long-context stability combination\n- 262,144-token context, 150,144 vocabulary, knowledge cutoff June 2026\n- BF16 weights about 80GB, with an official SGLang serving kit for B300\u002FH200\n\nThe sliding-window emphasis is a direct product of the coding-agent role: in codebase-search and fix loops most of the prefix is reusable, a 513-token window shrinks KV to almost nothing, and attention cost yields to expert bandwidth.\n\n## Community GGUFs (unsloth, spread by Oct 21)\n\n| Quant | File size |\n|---|---|\n| UD-IQ1_M | 11.4 GB |\n| UD-Q2_K_XL | 13.6 GB |\n| UD-Q3_K_XL | 18.5 GB |\n| **UD-Q4_K_XL** | **25.1 GB** |\n| UD-Q5_K_XL | 29.9 GB |\n| UD-Q6_K_XL | 36.3 GB |\n| Q8_0 | 42.3 GB |\n\nQ4_K_XL at 25.1GB means a single 32GB RTX 5090 can hold a 40B-class coding model — physically speaking; note the site's VRAM accounting uses a deliberately padded formula (KV\u002Fruntime headroom included), so this wave's runs landed mostly on ≥60GB-class hardware.\n\n## On-site coverage\n\n**Flint-40B-A2.5B** (id 92) is now cataloged with **22 new L2 records** (llama.cpp \u002F vLLM \u002F Ollama \u002F Strata), averaging 156.2 tok\u002Fs across the pool:\n\n| Hardware | Quant | Measured |\n|---|---|---|\n| RTX PRO 6000 96GB | Q3_K_S | 343.5 tok\u002Fs |\n| H100 | IQ2_M | 326.0 tok\u002Fs |\n| M5 Ultra | IQ1_S \u002F Q2_K | 324.8 \u002F 230.7 tok\u002Fs |\n| RTX Spark (116 GiB Linux pool) | Q4_K_M | 70-76 tok\u002Fs |\n| Strix Halo 128GB | Q4_K_S | 63.8 tok\u002Fs |\n\nAt 2.5B active it is the fastest model of this batch — MoE decode only reads the activated experts on the bandwidth bill, and 2.5B active runs an order of magnitude faster than a dense 40B.","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Frecursive-flint-40b-2026-10-22","2026-10-22T15:23:50.378Z"]