Research Vault

Recursive Flint (40B-A2.5B)

Properties1
tagshuggingface, text-generation, coding, agentic, instruct, moe, reasoning, tool-calling, local-inference, custom-architecture, us-lab

3 min read

Recursive Flint (40B-A2.5B)

One-line: Debut model from Recursive Superintelligence (Richard Socher; the $4B “self-improving AI” lab) — a 39.8B-param / ~2.5B-active MoE built from scratch for local agentic coding, Apache-2.0, trained end-to-end by their “Eureka Machine” automated-research loop; lab-reported SWE-bench Verified 74.2 / SWE-bench Pro v2 89.7 / Terminal-Bench 2.1 54.9, with a claim of >125 tok/s on an M4 Pro MacBook Pro and >150 tok/s on a 64GB RTX Spark.

⚠ All benchmark numbers below are lab-reported (model card + launch blog; agentic evals are average pass@1 over ≥5 attempts, 5-hour agent timeouts). No independent evaluation exists yet as of 2026-10-22 — Artificial Analysis and BenchLM both 404 on flint. The SWE-bench Pro v2 figure (89.7) is above every open-weight model on record (Kimi K3 88.2, GLM-5.3 84.3, reflection/Beam 76.5) for a 2.5B-active model, which warrants scepticism until reproduced; the card’s own caveat is that developer-reported comparison scores may not apply the same strict retrieval adjustment Flint’s Terminal-Bench score does.

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench v6—listed in the blog’s mean-score set, value not published
Aider polyglot—not published
SWE-bench Verified74.2lab — card + blog (mini-swe-agent, no internet)
SWE-bench Pro v289.7lab — card + blog (mini-swe-agent, no internet)
SWE-bench Multilingual—not published
Terminal-Bench 2.154.9lab — card + blog (Terminus 2 XML, internet on, retrieved solutions counted as failures)

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA—not published
Tool-use supportOpenAI-compatible tool calls, multi-turn, reasoning_content preservedcard (no BFCL/τ number published)

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published
GPQA Diamond—not published
MATH—not published
AIME 2025—not published
IFBench / IFEval—listed in the blog’s mean-score set, values not published

Efficiency

  • Active params: ~2.5B active per token (2.2B body + output head) / 39.8B total; 256 routed experts, 6 selected per token, plus one always-active shared expert
  • Context length: 262,144 documented (blog: supports up to 512k, recommends ≤256k); prompt and output share the window
  • VRAM (fp16 rough): ≈80 GB BF16 (15 safetensors shards, ~5.37 GB each) — one B300 or H200 per replica with the shipped SGLang kit. Realistic local path: the first-party flint-int4 / flint-int4-gguf / flint-nvfp4-gguf siblings (created the same day) or the Ollama library entry — the blog’s 125–150 tok/s local figures are on Apple Silicon / RTX Spark with their optimized engine
  • Runtime: ⚠ custom architecture (model_type: flint, FlintForCausalLM) and the card states Transformers modeling code is not included yet — day-0 local use means the repo’s SGLang serving kit (github.com/recursive-oss/rec) or a GGUF/llama.cpp path, not stock transformers
  • Knowledge cutoff stated as June 2026; vocab 150,144; temp 1.0 / top-p 0.95 / top-k 20

What makes it notable

First release from Recursive Superintelligence and the strongest “small-active” coding MoE on paper: 2.5B active params is laptop-class decode while the card claims SWE-bench Verified 74.2 and Terminal-Bench 2.1 54.9 — i.e. near reflection/Beam-class agentic coding at ~1/10th the memory of a 501B. The pitch is explicitly local-first (“keep your data on your computer, avoid API fees”, 125+ tok/s on an M4 Pro), and the training story is itself the product: the whole stack (architecture search, data, training recipes, inference kernels) was built by their automated “Eureka Machine” research loop in under four months. Caveats: every number is lab-reported and independently unverified; SWE-bench Pro v2 89.7 would be a new open-weight record and should be treated as an unverified claim; and the custom architecture + missing transformers code means it will not load in stock runtimes until community quants/ports land.

See also

Graph View