Recursive ships Flint: a 39.8B-A2.5B coding MoE hitting 89.7% on SWE-bench Pro v2

On 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.

Release evaluations (Recursive-measured)

Benchmark	Score
Terminal-Bench 2.1	54.9%
SWE-bench Pro v2	89.7%
SWE-bench Verified	74.2%

Methodology notes: temperature 1.0 / 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.

Architecture: a 49-layer, sliding-window-first design

• 49 layers = 2 dense + 47 MoE; 256 routed experts with 6 selected per token plus an always-active shared expert
• 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
• Per-head QK RMSNorm plus attention sinks — the usual long-context stability combination
• 262,144-token context, 150,144 vocabulary, knowledge cutoff June 2026
• BF16 weights about 80GB, with an official SGLang serving kit for B300/H200

The 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.

Community GGUFs (unsloth, spread by Oct 21)

Quant	File size
UD-IQ1_M	11.4 GB
UD-Q2_K_XL	13.6 GB
UD-Q3_K_XL	18.5 GB
UD-Q4_K_XL	25.1 GB
UD-Q5_K_XL	29.9 GB
UD-Q6_K_XL	36.3 GB
Q8_0	42.3 GB

Q4_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/runtime headroom included), so this wave's runs landed mostly on ≥60GB-class hardware.

On-site coverage

Flint-40B-A2.5B (id 92) is now cataloged with 22 new L2 records (llama.cpp / vLLM / Ollama / Strata), averaging 156.2 tok/s across the pool:

Hardware	Quant	Measured
RTX PRO 6000 96GB	Q3_K_S	343.5 tok/s
H100	IQ2_M	326.0 tok/s
M5 Ultra	IQ1_S / Q2_K	324.8 / 230.7 tok/s
RTX Spark (116 GiB Linux pool)	Q4_K_M	70-76 tok/s
Strix Halo 128GB	Q4_K_S	63.8 tok/s

At 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.