ManniX-ITA/Qwen3.6-27B-A3B-CoderX-MTP-GGUF
Text Generation • 26B • Updated • 3.48k • 2
ManniX PRO
ManniX-ITA
AI & ML interests
None yet
Recent Activity
updated a model about 4 hours ago
ManniX-ITA/Qwen3.6-27B-A3B-CoderX-MTP-GGUF updated a model about 4 hours ago
ManniX-ITA/Qwen3.6-27B-A3B-CoderX updated a collection 1 day ago
Qwen-3.8Organizations
None yet
updated 2
models about 4 hours ago
updated a
collection 1 day ago
updated a
model 1 day ago
published a
model 1 day ago
Best performing model of qwen3.6 family
🤗 1
27
#3 opened 10 days ago
by 
repla73
Suggestion: try pruning KAT-Coder-V2.5-Dev
👍 1
1
#1 opened 15 days ago
by 
mslxd
updated a
collection 3 days ago
updated 2
models 3 days ago
published 2
models 3 days ago
posted an update 3 days ago
Post
90
🚀 opencoti-llamafile 0.10.5-c7 — a serving API agents can trust
New cut of the opencoti single-file inference engine, rebased onto llamafile 0.10.5 / llama.cpp (mozilla-ai) with 190 additive patches (series published in the repo). Same zero-dependency APE: one executable for Linux, Windows, macOS & BSD.
What's new vs c6:
PolyKV admission is now atomic and enforced by default. Every admit is a reservation against the pool's KV budget — the optimistic check-then-book race under concurrent agent spawns is gone (was 0/4/5 nondeterministic refusals on the same binary; now deterministic booked/refused accounting). Pool-less servers get the same math via --admission-poolless (a clean refusal in 0.38s instead of an 11s stall), sequences can be partially evicted instead of dropped, and SWA models size and admit on the same budget model (--swa-seq-budget: −40 GiB KV measured).
llamafile as an agent skill. opencoti now ships an opt-in plugin + embedded skill that teaches AI agents to drive the local engine: launch it, check capacity before spawning sub-agents, fork shared-prefix pools from a live session so N agents share one cached system prompt, token-exact.
Agentic reliability: Gemma-4 tool-call argument bleed fixed. An un-closed string argument no longer swallows the brace, thought channel and the next tool call — contained at map time, grammar untouched.
Vulkan ships for the first time. x86_64 + Windows Vulkan side-load DSOs (AMD / iGPU / RADV) alongside the CUDA ones.
Per-cut side-load isolation. The DSO cache dir is namespaced by the full engine version, so different cuts no longer collide on extraction.
Binaries (Linux, Windows, Windows-GPU, aarch64, universal), side-load DSOs and docs: ManniX-ITA/opencoti-llamafile
New cut of the opencoti single-file inference engine, rebased onto llamafile 0.10.5 / llama.cpp (mozilla-ai) with 190 additive patches (series published in the repo). Same zero-dependency APE: one executable for Linux, Windows, macOS & BSD.
What's new vs c6:
PolyKV admission is now atomic and enforced by default. Every admit is a reservation against the pool's KV budget — the optimistic check-then-book race under concurrent agent spawns is gone (was 0/4/5 nondeterministic refusals on the same binary; now deterministic booked/refused accounting). Pool-less servers get the same math via --admission-poolless (a clean refusal in 0.38s instead of an 11s stall), sequences can be partially evicted instead of dropped, and SWA models size and admit on the same budget model (--swa-seq-budget: −40 GiB KV measured).
llamafile as an agent skill. opencoti now ships an opt-in plugin + embedded skill that teaches AI agents to drive the local engine: launch it, check capacity before spawning sub-agents, fork shared-prefix pools from a live session so N agents share one cached system prompt, token-exact.
Agentic reliability: Gemma-4 tool-call argument bleed fixed. An un-closed string argument no longer swallows the brace, thought channel and the next tool call — contained at map time, grammar untouched.
Vulkan ships for the first time. x86_64 + Windows Vulkan side-load DSOs (AMD / iGPU / RADV) alongside the CUDA ones.
Per-cut side-load isolation. The DSO cache dir is namespaced by the full engine version, so different cuts no longer collide on extraction.
Binaries (Linux, Windows, Windows-GPU, aarch64, universal), side-load DSOs and docs: ManniX-ITA/opencoti-llamafile
updated 3
models 4 days ago
posted an update 16 days ago
Post
162
🚀 Qwen3.6-27B-A3B-CoderX — the long-horizon sibling to A3B-Coder. Same 256→184 expert budget (~35B→27B, A3B active), different selection: our saliency map picks the keep-set, a REAP-style per-layer floor (p=24) protects the tail, and the 72 evicted experts per layer are folded DERN-style into the survivors instead of discarded. No fine-tuning, no distillation.
📊 Q6_K + imatrix, llama.cpp b9700, greedy, one pinned geometry per bench, same host — CoderX / A3B-Coder / unpruned 256e:
⚡ LiveCodeBench v6 (77q, 24k think) — 72.73 / 61.04 / 61.04 → +11.7pp over both
✅ HumanEval+ (164) — 96.95 / 95.12 / 93.90 → best of the three
🤝 MultiPL-E-100 (rs+java+js) — 88.67 / 89.00 / 91.00
⚠️ Read that last row honestly: a same-basis repeat of MultiPL-E moved 1.0pp on batch-scheduling nondeterminism alone. The 0.33pp CoderX↔Coder gap is INSIDE that band — a tie. The 2.33pp gap to the base is outside it and real. CoderX takes Rust (0.85 vs 0.81), gives up JS (0.92 vs 0.96).
🎯 Ships top-8, and that was measured, not assumed: MBPP-full 78.4 / 79.0 at top-8 vs 73.2 / 73.0 at top-10. Opposite call from A3B-Coder, which bakes top-10.
🧠 It thinks long — LCB median completion ~15.8k tokens vs ~2.2k for Coder. The length is where the win comes from; give it context headroom rather than clamping it.
🔬 Not measured yet: the canonical 9-bench. GPQA / MATH-500 / IFEval are deliberately NOT quoted — treat the non-code profile as unknown. Coder remains the one with a published 9-bench table.
📦 bf16 safetensors (text-only) · 19 GGUF tiers, EVERY K/I-quant imatrix-built and verified by reading quantize.imatrix.* back out of each uploaded file · Ollama 39 tags (19 text + 19 vision-<tier> + :latest). MTP in every tier — draft_num_predict 3 gives 190→252 tok/s (+33%) on an RTX 5080.
🔗 ManniX-ITA/Qwen3.6-27B-A3B-CoderX
🔗 ManniX-ITA/Qwen3.6-27B-A3B-CoderX-MTP-GGUF
🔗 https://ollama.com/mannix/qwen3.6-27b-a3b-coderx
📊 Q6_K + imatrix, llama.cpp b9700, greedy, one pinned geometry per bench, same host — CoderX / A3B-Coder / unpruned 256e:
⚡ LiveCodeBench v6 (77q, 24k think) — 72.73 / 61.04 / 61.04 → +11.7pp over both
✅ HumanEval+ (164) — 96.95 / 95.12 / 93.90 → best of the three
🤝 MultiPL-E-100 (rs+java+js) — 88.67 / 89.00 / 91.00
⚠️ Read that last row honestly: a same-basis repeat of MultiPL-E moved 1.0pp on batch-scheduling nondeterminism alone. The 0.33pp CoderX↔Coder gap is INSIDE that band — a tie. The 2.33pp gap to the base is outside it and real. CoderX takes Rust (0.85 vs 0.81), gives up JS (0.92 vs 0.96).
🎯 Ships top-8, and that was measured, not assumed: MBPP-full 78.4 / 79.0 at top-8 vs 73.2 / 73.0 at top-10. Opposite call from A3B-Coder, which bakes top-10.
🧠 It thinks long — LCB median completion ~15.8k tokens vs ~2.2k for Coder. The length is where the win comes from; give it context headroom rather than clamping it.
🔬 Not measured yet: the canonical 9-bench. GPQA / MATH-500 / IFEval are deliberately NOT quoted — treat the non-code profile as unknown. Coder remains the one with a published 9-bench table.
📦 bf16 safetensors (text-only) · 19 GGUF tiers, EVERY K/I-quant imatrix-built and verified by reading quantize.imatrix.* back out of each uploaded file · Ollama 39 tags (19 text + 19 vision-<tier> + :latest). MTP in every tier — draft_num_predict 3 gives 190→252 tok/s (+33%) on an RTX 5080.
🔗 ManniX-ITA/Qwen3.6-27B-A3B-CoderX
🔗 ManniX-ITA/Qwen3.6-27B-A3B-CoderX-MTP-GGUF
🔗 https://ollama.com/mannix/qwen3.6-27b-a3b-coderx
- 1 reply
updated a
collection 16 days ago